GAMUT MAPPING BASED ON NUMERICAL AND PERCEPTUAL MODELS

Abstract

Methods and systems herein provide for color conversion. Such color conversion includes gamut mapping by determining a color gamut of an imaging device and receiving color data having color values external to the color gamut of the imaging device. The gamut mapping proceeds by determining a color region external to the color gamut and mapping the color values within the color region to the color gamut. Thereafter, the area of the color region is increased and the color values within the increased color region are mapped to the color gamut.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The invention relates to the field of printing systems, and in particular, to methods and systems for color conversions.

2. Statement of the Problem

In color printing, displaying, and reproduction, the term gamut represents the set of colors that a color-reproduction device is physically able to generate. Every device that displays or reproduces an image, such as a printer, monitor, scanner, or digital camera, may have its own unique color gamut. When an image is transferred from one device to another, the color gamut of each device is examined.

The color gamut generally has two components that are considered when performing a color conversion. Those components include the gamut boundary and the number of colors that are realizable within the gamut boundary. The gamut boundary of a device represents the outermost extent of the device's capabilities in a reference color space. Because of quantization in color reproduction systems, such as those of color halftone printers, not all colors within a device's boundary are realizable. Moreover, the shape of the boundaries can vary dramatically.

When an input color space is larger than the gamut of an output color device, gamut mapping algorithms are applied. The gamut mapping process transforms a point in the source gamut to a realizable color inside the gamut of the output device. The form of this transformation can dramatically impact the quality of the reproduced images. As such, care is taken in the design and implementation of gamut mapping transformations.

Certain current gamut mapping algorithms map all out-of-gamut points directly to the destination gamut. For example, clipping algorithms clip out-of-gamut points to the destination gamut boundary. Scaling algorithms, on the other hand, scale an input color gamut to an output color gamut such that some of the out-of-gamut points are mapped inside the destination gamut while others are mapped to the boundary of the destination gamut.

The large variability in past color gamut mapping studies suggests that ideal gamut mapping depends on image content, preservation of perceived hue throughout color space and the extent of the gamut mismatch in various regions of color space. Thus, image dependent and regional-dependent gamut mappings have been preferred. One such gamut mapping includes a linear conversion function that interpolates color space source data to printer specific color space data by subdividing the source data into a plurality of subsets and interpolating the data for each subset using one or more linear conversion functions. A problem, however, exists with the linear conversion function as it does not generate an optimum interpolation for each subset. To compensate for sub-optimal interpolations of subsets, one or more of a plurality of conversion functions may be selected for each subset, optimizing the interpolation for each subset. The converted subsets, however, still do not optimally combine to form a target data set.

Perceptual gamut mapping is also used in color reproduction. Perceptual gamut mapping modifies both in-gamut and out-of-gamut colors from their colorimetric representation in order to provide a pleasing or perceptual appearance. The results of the perceptual gamut mapping normally depend on the input color range and the device color gamut. However, if the input color range is much larger than the actual output color range, as is normally the case, the reproduced colors have lower chroma and are less vivid. Accordingly, there is exists a need to map colors from the gamut of one device to the gamut of another device while ensuring that the colors remain aesthetically pleasing and vivid.

SUMMARY OF THE SOLUTION

Embodiments of the invention solve the above and other related problems with methods and associated systems by providing gamut mapping based on a numerical model that is guided by a perceptual color difference model. The gamut boundary is incrementally expanded into layers. For example, out-of-gamut CIELab colors in each layer may be converted to the CMY values using radial basis function (RBF) models generated from data in a previous layer. These CMY values are corrected by a perceptual color difference model. The conversion transforms an out-of-gamut color set to a set of colors around the boundary of the device gamut. While the out-of-gamut colors may vary continuously, the direction of the gamut mapping varies smoothly.

In one embodiment, a method of color conversion includes determining a color gamut of an imaging device and receiving image data having color values external to the color gamut of the imaging device. The method also includes determining a color region of the image data external to the color gamut and mapping the color values within the color region to the color gamut. The method also includes iteratively increasing an area of the color region and mapping the color values within the increased color region to the color gamut.

In another embodiment, a color conversion system includes an imaging device operable to generate image data having a plurality of color values and a processor operable to convert the color values of the image data to a color gamut of another imaging device. The processor determines a color region external to the color gamut, maps the color values within the color region to the color gamut, iteratively increases an area of the color region, and maps the color values within the increased color region to the color gamut.

In another embodiment, a printing system includes a printer and an imaging device operable to generate image data having a plurality of color values. The printing system also includes a printer controller operable to process the color values of the image data and map the color values to a color gamut of the printer by determining a color region external to the color gamut, mapping the color values within the color region to the color gamut, increasing an area of the color region, and mapping the color values within the increased color region to the color gamut.

In another embodiment, a method of color conversion includes determining a color gamut of an imaging device and receiving image data having color values external to the color gamut of the imaging device. The method also includes determining a color region of the image data external to the color gamut and mapping the color values within the color region to the color gamut via a numerical model of the color gamut. The method further includes mapping the mapped color values to the color region via a numerical model of the color region and comparing the mapped color values of the color region to original color values of the color region to determine whether the mapped color values correspond to the original color values.

The invention may include other exemplary embodiments described below.

DESCRIPTION OF THE DRAWINGS

The same reference number represents the same element or the same type of element on all drawings.

FIG. 1 is a block diagram of a color range gamut mapping system in one exemplary embodiment of the invention.

FIGS. 2 and 3 illustrate a color gamut and gamut mapping in one exemplary embodiment of the invention.

FIG. 4 illustrates the non uniformity of CIELab chroma.

FIG. 5 is a graphical illustration of the Minimax Characterization Theorem (MCT).

FIG. 6 is a graph of exemplary toner/ink values used in color conversion in one exemplary embodiment of the invention.

FIG. 7 is a flowchart illustrating a method of gamut mapping in one exemplary embodiment of the invention.

DETAILED DESCRIPTION OF THE INVENTION

FIGS. 1-7 and the following description depict specific exemplary embodiments of the present invention to teach those skilled in the art how to make and use the invention. For the purpose of this teaching, some conventional aspects of the invention have been simplified or omitted. Those skilled in the art will appreciate variations from these embodiments that fall within the scope of the present invention. Those skilled in the art will appreciate that the features described below can be combined in various ways to form multiple variations of the present invention. As a result, the invention is not limited to the specific embodiments described below, but only by the claims and their equivalents.

FIG. 1 is a block diagram of a color range gamut mapping system 100 in one exemplary embodiment of the invention. The color range gamut mapping system 100 includes at least two imaging devices having differing color gamuts. As mentioned, in color reproduction, a color gamut is a set of discrete colors that are used to represent the color capabilities of a particular device. Different imaging devices may use different color gamuts to represent those color capabilities. For example, a digital camera may use one color gamut to represent an actual image while a printer may use another color gamut to represent the image captured by the digital camera. In this context, digitizing a photograph and converting it to a different color space generally alters the color gamut in the sense that some of the colors in the original image of the digital camera are lost in the process.

To perform the gamut mapping, the gamut mapping system 100 may include a computer system 102 that communicatively couples between devices to convert the color gamut of one device to another. For example, the computer system 102 may be communicatively coupled to a printer 106 with an output color gamut and to one or more imaging devices, such as camera 116 and scanner 114, with input color gamuts. The computer system 102 may include a processor 110 and software instructions that direct the processor 110 to prepare an image captured by the imaging device for printing on the printer 106. In doing so, the computer system 102 may determine a color gamut of the captured image as well as a color gamut of the printer 106 and numerically map the color gamut of the captured image to that of the color gamut of the printer 106 through the use of numerical models of the color gamuts. The gamut mapping described herein provides for a more vivid range when applied to the printer medium 108 (e.g., paper).

Although shown and described as being configured within a computer system 102, the invention is not intended to be so limited. For example, the gamut mapping described herein may be implemented in software instructions that are operable within a variety of devices, such as the print controller of a large-scale printing system. Thus, the processor 110 is merely intended to illustrate any processor suitable for implementing the gamut mapping algorithms described herein. The gamut mapping algorithms also provide for the mapping of color values between the color gamuts of virtually any imaging device and not simply from an imaging device to a printer. For example, the gamut mapping algorithms described herein may be used from the gamut of the scanner 114 to the gamut of the computer monitor 112 or like device.

To illustrate the numerical gamut mapping, FIG. 2 shows a color gamut 201 of an output device such as the printer 106 and the color gamut 206 of an image, such as that produced by the digital camera 116. Initially, each color gamut is determined such that the color gamut 201 may be mapped within the color gamut of the image. That is, the color gamut 201 of the imaging device is generally smaller than the color gamut 206 of the image. Accordingly, colors from the image are mapped from the color gamut 206 to the color gamut 201 such that, when reproduced by the output device, the mapped colors appear comparable to that of the original image. In this regard, an exemplary color point 203 may be numerically mapped from the color gamut 206 to the color point 202 within the color gamut 201.

To map the color point 203 from the color gamut 206 to the color point 202, a numerical model is generated that provides for the mapping of regions of color points within the color gamut 206 to the color gamut 201. For example, the printer 106 may be a color laser printer that employs CMYK printing. The color gamut 206 of the image may be CIELab. CMYK and CIELab are known to those skilled in the art. A numerical model for this CIELab to CMYK mapping is generated using radial basis functions (RBF), which are described in greater detail below. In doing so, a layer 204 of color points surrounding the color gamut 201 and containing the color point 203 is defined relatively close to the boundary 207 of the color gamut 201. The color point 203 is then mapped to the color point 202 using the numerical CIELab to CMYK model. For that matter, each of the color points contained in the layer 204 is mapped to color points within the color gamut 201 using the numerical CIELab to CMYK model.

A numerical model for the CMYK to CIELab mapping may also be generated using RBFs. For example, once the layer 204 is mapped to the color gamut 201, the color point 202 may be mapped back to the color gamut 206 using a CMYK to CIELab model to determine whether the mapping is aesthetically appropriate. From there, the color point 202 should map to the color point 203 or proximate thereto meaning that the original mapping provided a sufficient approximation of the original color at color point 203.

To ensure that the mapping has a sufficient approximation, the newly mapped color point is compared to the original color point 203 using an RBF color conversion model and a perceptual color difference model, such as CIELab. If the newly mapped color point perceptually matches using the perceptual color difference model, then the mapping is deemed a success and mapping of color points outside the layer 204 may continue as illustrated in FIG. 3 with layer 205. Otherwise, the color point 203 is represented by the corresponding color point of the perceptual color difference model.

Before discussing the numerical models for CIELab to CMYK and CMYK to CIELab color conversions in detail, a discussion of color spaces is now presented to guide the reader through an understanding of color conversions and gamut mapping. Perceptual color spaces, such as CIEXYZ and CIELab, are visualized as three dimensional color spaces, where every color that humans can see is uniquely located. The location of any color in the space is determined by its color coordinates. In the CIEXYZ color space, the components are X, Y, and Z, where Y is luminance, X and Z do not correlate to color appearances. In the CIELab color space: the components are L*, a*, and b*, where L* is lightness of the color (e, g., L*=0 yields black and L*=100 indicates white), a* is the value between red and green (e.g., negative values indicate green and positive values indicate red), and b* is the value between yellow and blue (e.g., negative values indicate blue and positive values indicate yellow). The CIELab color space can also be described using cylindrical coordinates in the L*C*h color space where C*_ab is the chroma coordinate, and h is the hue coordinate. Below are the mathematical expressions to convert a* and b* into C*_ab and h:

$\begin{array}{cc}{C}_{\mathrm{ab}}^{*}=\sqrt{a^{*2}+b^{*2}}& \left(\mathrm{Eq}.1\right)\\ h=\tan \left(\frac{b^{*}}{a^{*}}\right)& \left(\mathrm{Eq}.2\right)\end{array}$

The relationship between CIEXYZ and CIELab may be described using the following equations:

$\begin{array}{cc}{L}^{*}=116f\left(\frac{Y}{Y_n}\right)-16,& \left(\mathrm{Eq}.3\right)\\ a^{*}=500\left(f\left(\frac{X}{X_n}\right)-f\left(\frac{Y}{Y_n}\right)\right),& \left(\mathrm{Eq}.4\right)\\ b^{*}=200\left(f\left(\frac{Y}{Y_n}\right)-f\left(\frac{Z}{Z_n}\right)\right),& \left(\mathrm{Eq}.5\right)\end{array}$

where X_n, Y_n, Z_n are the tristimulus values of a reference white and generally referred to as the white point, described as

$\begin{array}{cc}f\left(x\right)=\{\begin{array}{cc}{x}^{\frac{1}{3}}& x>0.008856\\ 7.787x+\frac{16}{116},& x\le 0.008856.\end{array}& \left(\mathrm{Eq}.6\right)\end{array}$

The color difference in either CIEXYZ or CIELab is generally calculated using the Euclidean distance. Though the CIEXYZ and CIELab color spaces are perceptual color spaces, they are not perceptually uniform color spaces as the Euclidean distance in the space does not correspond to perceptual distance. For example, the colors that have the same Euclidean distance with respect to an anchor color are on a circle. However, these circles are transformed to ellipsoids when the human perceptual distances are measured. The major axis of the ellipsoid represents the higher tolerance of the color difference and the minor axis of the ellipsoid represents the lower tolerance for the color difference. The CIELab color space is a perceptually more uniform color space than the CIEXYZ color space. Thus, the CIELab color space is the color space used most often for color-difference evaluations.

In the CIELab color space, the magnitude of the perceptual color difference generally depends upon the color location and the changing direction in chroma and hue. CIELab increasingly overstates the magnitudes of perceived chroma differences. For example, the human visual system (HVS) is sensitive to the change of chroma in the neutral color area. The HVS is insensitive to the change of the chroma in a highly saturated color area. FIG. 4 illustrates such change with the ellipsoid size increases 301.

The CIELab color space is also non-uniform regarding hue angle in that the thresholds of visual tolerances are a function of hue angle. If the non-uniformity of the CIELab color space in chroma and hue is examined from another perspective, CIELab colors have different characteristics at different locations. The change of these characteristics is generally continuous. Color difference models correlate the perceptual distance with the Euclidean distance in the CIElab color space based on vast amounts of color difference experimental data.

There are several CIE perceptual-based color difference models. For example, CIEDE2000 calculates the perceptual difference by weighted lightness difference, chroma difference, hue difference, and the interaction between the chroma and the hue difference. However, the CIEDE2000 was developed for small color differences and does not perform well for large color differences. The empirical weighted color differences of lightness, chroma, and hue are often used for CIEDE2000, with higher weighting factors for hue and lightness differences and lower weighting factor for chroma differences. The color difference calculated using the Euclidean distance in CIELab is called deltaE, (also referred as ΔE, or dE_ab.) dE₀₀ represents the color difference calculated using CIEDE2000.

A typical imaging system consists of a source device (e.g., scanner 114, digital camera 116, etc), a display device (e.g., monitor 112) and a destination device (e.g., printer 106). Each of these devices produces and interprets colors differently. This creates a significant problem in terms of color accuracy and consistency in color reproduction. For example, a customer's banking statement may contain several objects that originated from different devices and software, such as a personal photo taken by a digital camera, a car advertisement scanned from a picture, the background graphics created by a graphic package, etc. It is difficult to predict what these colors look like when displayed on different monitors or printed from different printers. The color management described herein provides for more consistent and predictable colors throughout all parts of the color reproduction chain.

In general, there are two types of color management systems among color devices: additive systems such as the monitor 112 and subtractive systems such as the printer 106. In additive systems, colors are transmitted lights. Red, green and blue lights are referred as the additive primary colors. When used in various degrees of intensity and variation, they create all other colors of light. When superimposed equally, they create gray. Primary colors of subtractive systems like printers may be cyan, magenta, yellow, and black that are used together to effectively create a multitude of other colors based on the subtractive color theory.

The color conversions for an additive system are normally between the RGB color space and CIEXYZ color space. These conversions are relatively straightforward since the gamma corrected RGB primaries are linearly related to XYZ values in the CIEXYZ space for most devices. An example of such conversion is the conversion for the RGB colors described as follows:

$\begin{array}{cc}\left(\begin{array}{c}X\\ Y\\ Z\end{array}\right)=\left(\begin{array}{ccc}0.4124& 0.3576& 0.1805\\ 0.2126& 0.7152& 0.0722\\ 0.0193& 0.1192& 0.9505\end{array}\right)\left(\begin{array}{c}g\left(R_{\mathrm{srgb}}\right)\\ g\left(G_{\mathrm{srgb}}\right)\\ g\left(B_{\mathrm{srgb}}\right)\end{array}\right),& \left(\mathrm{Eq}.7\right)\end{array}$

where

$\begin{array}{cc}g\left(k\right)=\{\begin{array}{cc}{\left(\frac{K+\alpha }{1+\alpha }\right)}^{\gamma },& K>0.4045\\ \frac{K}{12.92},& \mathrm{otherwise},\end{array}\gamma =2.2,\mathrm{and}\alpha =0.055.& \left(\mathrm{Eq}.8\right)\end{array}$

As is known, color printers may be toner-based or ink-based. Most of the printers are only able to produce a limited numbers of gray levels( e.g., a binary printer produces two levels (i.e., 0 and 1) 2-bit printers produce levels of 0 1, 2, and 3). In order to produce continuous tone imagery that contains an infinite range of colors or grays, a reprographic technique called halftoning is applied to create the illusion of continuous tone images through the use of dot arrangements and dots of varying size. The combination of the printer halftone design and the specific toner/ink selection determine the number of colors that a printer is physically able to produce which is called its gamut. The gamut is measured in the CIELab color space.

Printer color conversions depend the toner/ink selections and the halftone design. The conversions are generally performed between a device-dependent color space (CMYK) and a device-independent color space (CIELab). Often the format of the color conversion is an equally spaced lookup table, the values of which are either 1-byte or 2-byte integers.

In generating a printer color conversion, CMYK color patches whose values span the CMYK color space are printed. Each component value of CMYK is normally a 1-byte integer. The printed patches are measured using a spectral photometer which measures the spectral reflectance of each patch under a standard illuminant. The software in the spectral photometer calculates the tristimulus values and converts these values to CIELab values. The CIELab values are floating point numbers. Followed by the implementation of a data modeling algorithm and a gamut mapping algorithm, the color conversions LUTs are generated using this set of data. The CMYK and the CIELab values are normalized to the range of 0 to 1 for these calculations. Both CMYK to CIELab and CIELab to CMYK conversions may be generated for a color printer. The CMYK to CIELab color conversion is used when an image generated in this CMYK color space is to display on a monitor or to print on a different color printer. In other words, this conversion is used when the printer CMYK color space is used as a source color space. The CIELab to CMYK color conversion is used when the CIELab color space is the source color space.

For many color printers, the relationship between CMYK color space and the CIELab color space is highly nonlinear due to the interactions of cyan, magenta, yellow, and black planes in the printer subtractive system, thereby requiring complex functions. Another complication to the printer color conversion is due to the halftone design. Most halftone techniques are capable of creating color gray levels continuously. However, an abrupt change of two adjacent levels is a weakness in many halftone designs. Thus, the color conversion between CMYK and CIELab may be continuous but not differentiable.

The CMYK to CIELab color conversion is neither injective nor subjective. For example, every color in the CMYK color space may be mapped to a color in the CIELab color space. However, different CMYK values may also be mapped to the same CIELab value. So the CMYK to CIELab color conversion is not infective. The range of CIELab color space is much larger than the range of CMYK color space. That is, for many CIELab values, there are no CMYK values such that f(C, M, Y, K)→(L, a, b). Thus the CMYK to CIELab color conversion is not subjective.

Additionally, the CIELab to CMYK conversion is from R3 to R4. Thus, only one degree of freedom in this conversion exists and one CIELab value may be mapped to several CMYK values. On the other hand, because of the smaller gamut of the color printer, not every CIELab value can be mapped to a CMYK value. Since no function exists for the CIELab to CMYK color conversion, the conversion is achieved by converting the CIELab values to CMY values first and a special rule is created for adding the black toner to each CMY value.

The CIELab to CMY color conversion within a printer gamut is both injective and subjective. That is, virtually every CIELab value inside the printer color gamut can be mapped to a unique color in the CMY color space. So the CIELab to CMY color conversion is injective within color gamut. Additionally, for CMY values, there exists a CIELab value such that g (L, a, b)→(C, M, Y) where the CIELab to CMY color conversion is subjective. Out-of gamut CIELab values generally cannot be mapped to a CMY value. Therefore, no functions exist for the CIELab to CMY conversion outside the printer gamut and color conversion between the CMYK and CIELab color spaces is achieved by approximation.

Understanding the sources of error is important for the design of an approximation model. In addition, it is important to understand how these errors interfere with human perceptions for the design of the printer color conversion model. There are four significant sources of error in the color conversion process. These errors are normally measured with the Euclidean distance in the CIELab color space, ΔE, and include:

• 1. measurement errors from the instrument, which are normally smaller than 1 ΔE;
• 2. errors due to the machine reliability and repeatability, which for a well-calibrated printer are generally smaller than 2 units of ΔE;
• 3. errors due to paper roughness and non-uniform paper fibers and fillers, which are typically smaller than 1 unit of ΔE; and
• 4. errors from the model of the color conversion, which average around 3-4 units of ΔE with a maximum error between about 10 to 20 units of ΔE.

Because human perceptual tolerance varies from location to location in the CIELab color space, care is taken in the approximation algorithms in different color areas. For colors in neutral areas, the visual color difference tolerance is low (around 1 ΔE). The interpolation errors in these regions are generally small. If the accumulated errors from the first three categories are added to the modeling data, the effect of the outliers in the modeling set should be minimized. For the medium colors, with the increase of the chroma values in these areas, the perceptual tolerance increases. Though the outliers still exist, they play a much less important role in the color conversion. It is appropriate to design an approximation model to minimize the least squares error. For colors in the highly saturated color regions, interpolation errors are relatively high. The approximation algorithm is used to minimize the maximum error in the interpolation.

In general, color conversion approximation algorithms aim at minimizing least square errors. Due to the sources of error described above, emphasis is placed on the L_p error criteria, in particular the L₁ and L_∞ error for a robust neutral area color conversion and the reduction of maximum error in the saturated color regions. One approximation model includes the natural neighbor interpolation method, which is based on the Voronoi region. A Voronoi region V_i is a set of points for which a center c_i is the nearest center vector, as follows:

V _i ={x ∈ R ⁿ :i=argmin_j∈I∥x−c_j∥},   (Eq. 9)

where I is the set of center indices. This interpolation method is a local region-based weighting method generally provides better underlying printer characteristics.

Global interpolation methods may also include the inverse distance weighting method and the RBF method. For the inverse distance weighting method, data points are used to interpolate a value by assuming that interpolated values are affected more by nearby points and less by more distant points. The assumption is not accurate for colors in many regions. Thus, this method has yielded relatively poor color conversion results. RBF data fitting equipped with an L_p error-criterion, on the other hand, improves color conversion. The embodiments described herein also take into account perceptual color difference tolerance. As mentioned, perceptual color difference tolerances generally vary over different CIELab color regions. Taking this into account, the RBF data fitting technique has the advantage of minimizing color differences in different areas with respect to human perceptions or yield results.

Although black toner/ink is one of the primaries in a color printer, the black toner/ink actually serves as a “secondary” toner in color printing. Because the mix of cyan, magenta, and yellow toner/ink produces a “muddy brown” instead of black, black toner/ink is added in to make a more rich black and thereby make other colors richer and darker. Black toner/ink also typically costs less than cyan, magenta, and yellow toner/ink making the embodiments herein more advantageous from a cost perspective.

Printer CIELab to CMYK color conversion maps from R3 to R4. Additionally, CIELab is mapped to the CMY color space using an interpolation technique, then black toner/ink denoted by K, is added. The amount of cyan, magenta, yellow toner/ink may be adjusted by applying the under color removal (UCR) and the under color addition (UCA) corrections. The typical way of applying UCR is to remove equal amounts of yellow, magenta, and cyan and replace them with the same amount of black. Gray component replacement (GCR) is the result of UCR followed by UCA. However, colors resulting from UCR are generally dark, less saturated, and hue shifted. UCA may be applied to improve the weakness of UCR. Small portions of the three colors of cyan, magenta, and yellow may be added back to these areas. The UCA process is empirical and relatively processor intensive due to the mapping from a lower dimensional space to a higher dimensional space (e.g., R3 to R4) and the lack of a high fidelity model describing the nonlinear behavior of each toner. UCA is normally achieved by printing many patches and choosing the best results. In practice, the UCR and UCA correction schemes introduce errors in the color conversion process.

Because the CIELab to CMYK conversion is a map from R3 to R4, a set of CMYK values corresponding to the same CIELab value may be established. In this regard, g denotes a mapping from the CIELab values to the CMYK values such that

∃{(C,M,Y,K)⁽ⁱ⁾,i=1, 2, . . . k},s.t., g⁻¹(C,M,Y,K)⁽ⁱ⁾=(L,a,b),   (Eq. 10)

where, i=1; 2; . . . ; k. These level sets of CMYK values may be found through an optimization process if approximation models are developed for the CMYK to CIELab conversion and the CIELab to CMY conversion.

As mentioned, gamut mismatch is problematic in color reproduction systems since each device has a different gamut. In general input devices such as digital cameras, scanners, etc., generate more colors than the output devices such as color printers. When the input color space is larger than the gamut of the output color device, gamut mapping algorithms are applied. The gamut mapping process transforms a point in the source gamut to a realizable color inside the gamut of the output device (e.g., a destination gamut). The form of this transformation can impact the quality of the reproduced images, especially print images. For high-end digital printers, the gamut is relatively small. The maximum error of color conversions normally derives from mapping an outside-gamut color to an inside-gamut color mapping.

In the gamut mapping process, a device color gamut boundary may be initially identified. The calculated boundaries are normally connected lines. Previous gamut mapping algorithms mapped out-of-gamut points directly to the destination gamut. The most typical gamut mappings include clipping and scaling algorithms. Clipping algorithms clip out-of-gamut points to the destination gamut boundary. Scaling algorithms, on the other hand, scale the input color gamut to output color gamut such that some out-of-gamut points are mapped inside the destination gamut while others are mapped to the boundary of the destination gamut. The mapping directions may include a choice of lines along which the mapping is applied (e.g., a line changing in chroma at constant lightness and hue) that are often decided based on visual experiments. The mapping directions may also be along the lines of constant perceptual attribute predictors (e.g., constant lightness and hue, constant saturation and hue), along the lines towards a single center-of-gravity (e.g., the center is the compression towards L*=50 on lightness axis), along the lines towards variable center-of-gravity (e.g., the cusp is the compression towards lightness of cusp on lightness axis), or along the lines towards the nearest color in the reproduction gamut (e.g. as is the case with minimum ΔE clipping).

The large variability in past color gamut mapping suggests that no universal gamut mapping algorithm works for every device and image. Ideal gamut mapping generally depends on image content, preservation of perceived hue throughout color space, and the extent of the gamut mismatch in various regions of color space. Relationships among color data developed by a numerical model have not previously been considered in the gamut mapping process. The underlying relationship among color data may more accurately retain or rationally transform after gamut mapping. Such is achieved herein using a numerical model with guidance from the perceptual model.

The numerical model quantitatively compares the color difference between the original color and the printed color. The original colors are the CIELab colors. These colors are converted with the printer color conversion to the CMYK colors. The CMYK colors are sent to the printer to print and then measured by spectral photometers. The numerical method tests the approximation errors inside color gamut and gamut mapping results out-side the color gamut. The mean error and the maximum error may be calculated using either dE_ab or dE₀₀ for the color conversion inside the printer gamut. The maximum conversion error is often weighted more than the mean error in the evaluation because the color difference is more objectionable than the average color difference.

The perceptual model, on the other hand, correlates human perceptions with physical stimuli. A set of test images including pictorial images, vector graphics, and text samples are printed from the printer. The colors in the images include highlight colors, low-light colors, saturated colors, and out-of-gamut colors. Observers look at these images and make perceptual judgments based on the selected psychophysics method, such as the paired comparison method. On such experiments, observers are asked to compare an image pair converted by different color conversion methods. The original photos are normally presented as the reference. Typical questions are “Which one do you like better?”, “Which one is more similar to the original image?”, etc. The data is analyzed and converted to perceptual values using Thurstone's law of comparative judgment.

To assist the reader in the development of the numerical model, this disclosure broken into certain sections. Specifically, the sections are entitled “FITTING NON-LINEAR SCATTERED DATA”, “OVERDETERMINED LINEAR SYSTEMS IN THE L_∞ SENSE”, “COLOR CONVERSION USING THE L₁ NORM APPROXIMATION”, “ONE DIMENSIONAL COLOR MANIFOLDS IN THE CMYK COLOR SPACE”, “COLOR CONVERSION WITH INK LIMITATIONS”, and “GAMUT MAPPING”. The section entitled “FITTING NON-LINEAR SCATTERED DATA” provides the reader with a detailed description of solving printer color conversion problems using RBFs and data fitting techniques adapted to the L_∞ and the L₁ error criteria. The section entitled “OVERDETERMINED LINEAR SYSTEMS IN THE L_∞ SENSE” presents the reader with the Barrodale and Phillips dual algorithm (the “BP” algorithm) and the Bartels, Conn, and L₁ primal algorithm (the “BC” algorithm) for converting relatively large sparse problems to linear programming problems. The section entitled “COLOR CONVERSION USING THE L₁ NORM APPROXIMATION” illustrates a general L₁ approximation for overdetermined linear systems, the BR algorithm for solving the primal problem, and the L₁ approximation as it is used in color conversion. The section entitled “ONE DIMENSIONAL COLOR MANIFOLDS IN THE CMYK COLOR SPACE” introduces the reader to an optimized GCR algorithm with relatively high quality toner/ink selection. The section entitled “COLOR CONVERSION WITH INK LIMITATIONS” discusses toner/ink limitations implemented as a constraint in the linear programming problem formulations. The section entitled “GAMUT MAPPING” introduces the reader to the concept of numerical gamut mapping based on the RBFs discussed above and enhanced by the perceptual gamut mapping.

Fitting Non-Linear Scattered Data

Problems with the interpolation techniques are overcome through the use of RBFs that include data fitting techniques adapted to L_∞ and L₁ error criteria. Additionally, a perceptual color model is used to guide the implementation of the RBFs. For example, in order to solve nonlinear approximation problem using RBFs, a clustering algorithm is employed using the combination of the Linde-Buzo-Gray (LBG) algorithm and the covariance matrix-based algorithm. RBFs may then be selected and parameters thereof may be optimized. Approximation algorithms are generated for the L_p norm (emphasis on the L_∞, L₂, and L₁ norm) for the neutral, intermediate, and saturated color areas. A K-fold cross validation then refines the model parameters and an optimized overlapping area between two different models is determined.

Theoretically, a color printer is capable of producing at least 4 bytes of colors (1 byte each of cyan, magenta, yellow, and black). A subset of these CMYK colors and their associated CIELab values may be used to generate a color conversion model. Because of the underlying complicated characteristics of a color printer, a large CMYK data set and the associated CIELab values (often above 2000 data points) are selected to generate the color conversion model. The data selection depends in part on the halftone design which occasionally yields a nonlinear response. The data selection also depends on the perceptual color difference tolerance as the perceptual color difference tolerances vary over different CIELab color regions. Therefore, different color conversion models should be generated for different regions based on the perceptual color difference tolerance. The continuity and smoothness of these color conversion models are important to the overall quality of the color conversion. RBFs provide the flexibility of applying to both the interpolation and approximation data fitting problems. RBFs are useful to interpolate scattered data as the associated system of linear equations is invertible under relatively mild conditions on the locations of the data points.

Generally, RBFs are useful for approximating an unknown function from the data in the form

$\begin{array}{cc}f\left(x\right)=\mathrm{Ax}+a_0+\sum _{i=1}^Nw_i\phi \left(x-c_i\right),& \left(\mathrm{Eq}.11\right)\end{array}$

where f(x) is assumed to be continuous and differentiable over a compact domain of x ∈ Rⁿ and f(x)∈ R. The term Ax represents the linear component of the mapping and the offset a₀ allows for the mean of the data to be non-zero. The nonlinear portion of the map is represented by the superposition of the nonlinear function φ weighted by vectors w_i.

The RBF has several important features. For example, the weight parameters may be determined using linear models, the RBFs may be either local or global and the locations of the basis functions may be adapted using different clustering routines. For color conversion, the Voronoi region, for example, may provide relatively good clustering because it can incorporate the color difference model to characterize each cluster with its unique characteristics. Also, RBFs generally do not require data to lie on any sort of regular grid and the function f(x) is continuous and differentiable.

Traditionally, the weights w_i were determined by minimizing the mean-square error. However, due to the unique error characteristics in color conversion, the algorithms for solving the RBFs' weights using L_p norm emphasizes L_∞ norm, L₂ norm, and L₁ norm. Accordingly, the approach described herein includes choosing a training/testing data set for color conversion, clustering the color data, determining the centers optimizing the number of centers, and optimizing RBFs and parameter selections. Then, the L_p norm selections (specifically, the L_∞, L₂, and L₁ norms) may be optimized and the RBFs combined with linear programming for the L_∞ norm and L₁ norm approximation employing the duality theorem. Thereafter, the perceptual color model may be employed to improve the qualitative result.

The color conversion model is built based on the availability of a data set of input-output pairs, such as CMYK vs. CIELab. There are many ways to select a set of training data. For example, the CMYK colors in regular grids may be selected or the CMYK colors based on the halftone curves may be selected. In any case, the CMYK data is printed on a color printer and measured in the CIELab color space. For the CMYK to CIELab conversion, these CIELab values are random within the printer gamut and can be expressed for f: R4→R3 in the following form:

{x⁽ⁱ⁾=(C,M,Y,K)⁽ⁱ⁾ ∈ R4, i=1,2, . . . ,m}  (Eq. 12)

{y⁽ⁱ⁾=(L,a,b)⁽ⁱ⁾ ∈ R3,i=1,2, . . . ,m},   (Eq. 13)

where m is the number of color pairs in the data set. For the CIELab to CMYK conversion, g: R4^→R3

{x⁽ⁱ⁾=(L,a,b)⁽ⁱ⁾ ∈ R3, i=1,2, . . . ,m}  (Eq. 14)

{y⁽ⁱ⁾=(C,M,Y,K)⁽ⁱ⁾ ∈ R4,i=1,2, . . . ,m}.   (Eq. 15)

The printer color conversions described herein include both CMYK to CIELab and the CIELab to CMYK color conversions. The CIELab to CMYK conversion is achieved by converting the CIELab values to the CMY values first and then adding black to each CMY value via the GCR process. Therefore, the CMYK set should have enough C,M,Y, and K values. Generally, the grid points of cyan, magenta, yellow, and black components are used to represent the curve change of each component created by the halftone process. This means that the grids point may not be equally spaced.

Data clustering is used to identify the locations where basis functions should be placed in the model. In general, RBF algorithms focus on placing functions such that the mean square error is minimized. Due to the uniqueness of perceptual color characteristics, the location of the basis function should be placed in the local regions which have similar properties in terms of human visual tolerances. As mentioned, the CMYK values are device-specific values and do not directly relate to human perception. In the construction of the mapping g: R4^→R3and f: R4^→R3, clustering is performed on the CIELab data so the characteristics of the color data can be described by each cluster. The visual tolerances of color data may be determined using a color difference model. The most recent color difference model CIEDE2000 is used as the distance metric for each color relative to the center color because the distances are designed to be small.

The clustering algorithm also combines the Voronoi-based clustering algorithm with a subset selection algorithm. Because the CIELab data are random, a random center selection algorithm may be used as the first step for clustering/center selections. The clustering algorithms are then used to identify the clusters and to refine the center of each cluster. There are several efficient clustering algorithms available that may be used including “k-means,” LBG, and topology-preserving mappings. The algorithm generally proceeds by selecting an initial number N_c of centers randomly from the m data points in the training set and, for each center, computing the Voronoi set (i.e., all points in the data closest in the perceptual color space to this center relative to other centers). Then, the centers as the mean of the points in the Voronoi set are updated. This is repeated until the CIELab values of these centers no longer change. The clustering algorithm described above can be effectively combined with a subset selection algorithm, such as the orthogonal least squares (OLS) algorithm or the covariance matrix algorithm of commonly owned and co pending U.S. Patent Application Publication No. 2007/009796, the entire contents of which are incorporated by reference, to eliminate unnecessary clusters.

The OLS algorithm identifies which of the N centers are most useful in the RBF model and indicates which centers may be deleted from the model and at what expense to accuracy expressed in color difference. The computation of the weight parameters in Eq. 11 is an over-determined least squares problem. Thus, a solution is sought to the set of inconsistent equations,

y=Φw,   (Eq. 16)

where each column of the matrix Φ is associated with a single center and f does not actually reside in the column space of Φ. In general, a quality function is associated with a center. In this case, the value of a column ø_i of Φ is measured to solve the least squares. One measure is the cosine of the angle between f and each column of Φ, such as

$\begin{array}{cc}{v}_i={\left(\frac{f^T{\phi }^{\left(i\right)}}{f{\phi }^{\left(i\right)}}\right)}^2.& \left(\mathrm{Eq}.17\right)\end{array}$

In other words, the center is the one for which v_i is largest (i.e., where the angle is smallest). This solution is indexed by i* meaning ø_i* is the solution to optimization. Once the column and hence the RBF center of Φ has been determined, the next best center can be computed by projecting the remaining columns along ø_i*. This approach may be iterated to obtain a reduced subset of centers that has acceptable modeling accuracy. This algorithm may therefore proceeds as follows:

$\begin{array}{cc}\mathrm{Let}k=1,2,\dots ,N_c& \\ \mathrm{Step}1:k=1& \\ i_1=\text{arg}{\max \left(\frac{f^T{\phi }^{\left(i\right)}}{f{\phi }^{\left(i\right)}}\right)}^2& \left(\mathrm{Eq}.18\right)\\ q^{\left(1\right)}={\phi }^{\left(i_1\right)}& \left(\mathrm{Eq}.19\right)\\ \mathrm{Step}k=1& \\ {\alpha }_{\mathrm{jk}}^{\left(i\right)}=\frac{q_j^T{\phi }_i}{q_j^Tq_j},1\le j<k& \left(\mathrm{Eq}.20\right)\\ q_k^{\left(i\right)}={\phi }_i-\sum _{j=i}^{k-1}{\alpha }_{\mathrm{jk}}^{\left(i\right)}q_j& \left(\mathrm{Eq}.21\right)\\ i_k=\text{arg}{\max \left(\frac{f^Tq_k^{\left(i\right)}}{fq^{\left(i\right)}}\right)}^2\mathrm{where}i_k\le i\le N_c,i\ne i_1,\dots i\ne i_{k-1},\mathrm{and}& \left(\mathrm{Eq}.22\right)\\ q_k=q_k^{i_k}={\phi }_{i_k}-\sum _{j=1}^{k-1}{\alpha }_{\mathrm{jk}}q_k.& \left(\mathrm{Eq}.23\right)\end{array}$

The algorithm ends at the Mth step when

$\begin{array}{cc}1-\sum _{j=1}^M\left({\left(\frac{f^Tq_j}{fq_j}\right)}^2\right)<\rho ,\mathrm{where}0<\rho <1& \left(\mathrm{Eq}.24\right)\end{array}$

A cluster can also be described by its mean vector and variance-covariance matrix. The mean vector is the center of the cluster and the variance-covariance matrix consists of the variances of the color data points along the main diagonal and the covariance between each pair of color components in the other matrix positions. For example, if a cluster is in the CMYK color space, the covariance of this cluster provides a measure of the strength of the correlation between any two color components. Let the center CMYK value of a cluster be x=( C, M, Y, K) (i.e., x₁= C, x₂= M, x₃= Y, x₄= K) and let m_c denote the number of data points in a cluster. The formula for computing each element COV_jj′ in a 4×4 CMYK covariance matrix for a cluster is:

$\begin{array}{cc}{\mathrm{COV}}_{{\mathrm{jj}}^{\prime }}=\sum _{i=j}^{\mathrm{mc}}\left[\left(x_j^i-{\stackrel{\_}{x}}_j\right)\left(x_{j^{\prime }}^i-{\stackrel{\_}{x}}_{j^{\prime }}\right)\right]/\left(m_c-1\right),& \left(\mathrm{Eq}.25\right)\end{array}$

where j; j′=1, . . . ,4. When the rank of the covariance matrix is less than the number of the input color components, the correlations are not all independent of each other. Also, the number of the data points in the cluster is less than the number of color components.

A Gaussian function which can be expressed as the function of the covariance matrix is not selected if a non-full rank cluster exists. Thus, the clustering algorithm proceeds by calculating the covariance matrix of each cluster and, if the rank of the covariance matrix of a cluster is less than the number of input color components, the center of this cluster is deleted and the data points in the cluster are moved to nearby clusters. Thereafter, the center of each cluster is updated and the LBG algorithm is updated to refine each cluster. This is repeated until every cluster has a full rank covariance matrix.

The radial basis functions themselves are generally selected so that they satisfy an invertibility condition on the square interpolation problem for the weights. These functions have expansions, such as those given by Eq. 11, that represent continuous functions over compact domains. The location of these functions is stipulated by the vector centers {ck} generally scattered over the domain in a manner that reflects the distribution of the data. Both theoretical investigation and practical results suggest that the choice of the nonlinearity ø(.) is not critical to the performance of the RBF. However, for the color conversions from CMYK to CIELab, and CIELab to CMYK, the multiquadric function

φ(x)=√{square root over (r²+x²)}  (Eq. 26)

is generally more accurate. Other functions that may be used include the Gaussian

$\begin{array}{cc}\phi \left(x\right)=\exp \left(\frac{-x^2}{r^2}\right)& \left(\mathrm{Eq}.27\right)\end{array}$

and the inverse multiquadratic function

$\begin{array}{cc}\phi \left(x\right)=\frac{1}{\sqrt{r^2+x^2}}.& \left(\mathrm{Eq}.28\right)\end{array}$

Regardless of the function used, the radius r in Gaussian, multiquadratic, and inverse multiquadratic is optimized. Generally, the multidimensional Newton's method is an efficient optimization method for 2-norm least square settings. In this case, the objective function aims at minimizing the mean error. For example, using the color conversion from CMYK to CIELab:R4→R3,

$\begin{array}{cc}{f}_L\left(x_i,r_L\right)=A_Lx_i+a_{L0}+\sum _{j=1}^nw_{\mathrm{Lj}}{\phi }_L\left(x_i-c_j,r_L\right)& \left(\mathrm{Eq}.29\right)\\ f_a\left(x_i,r_a\right)=A_ax_i+a_{a0}+\sum _{j=1}^nw_{\mathrm{aj}}{\phi }_a\left(x_i-c_j,r_a\right)& \left(\mathrm{Eq}.30\right)\\ f_b\left(x_i,r_b\right)=A_Lx_i+a_{b0}+\sum _{j=1}^nw_{\mathrm{bj}}{\phi }_b\left(x_i-c_j,r_b\right)& \left(\mathrm{Eq}.31\right)\end{array}$

The problem formulation is thus:

$\begin{array}{cc}E\left(r_L,r_a,r_b\right)=\sum _{i=1}^m[\begin{array}{c}{f_L\left(x_i,r_L\right)-y_1^i)}^2+\\ {{f_a\left(x_i,r_a\right)-y_2^i)}^2+f_b\left(x_i,r_b\right)-y_3^i)}^2\end{array}]& \left(\mathrm{Eq}.32\right)\\ E_{\min }=E\left(r_L,r_a,r_b\right)\mathrm{Then}\mathrm{for}\mathrm{each}\mathrm{iteration},\mathrm{let}& \left(\mathrm{Eq}.33\right)\\ g^{\left(k\right)}=\left(\begin{array}{c}\partial E/r_L\\ \partial E/r_a\\ \partial E/r_b\end{array}\right)& \left(\mathrm{Eq}.34\right)\\ F\left(k\right)=\left(\begin{array}{c}\frac{{\partial }^2E}{{\partial }^2r_L}\frac{{\partial }^2E}{\partial r_L\partial r_a}\frac{{\partial }^2E}{\partial r_L\partial r_b}\\ \frac{{\partial }^2E}{\partial r_a\partial r_L}\frac{{\partial }^2E}{{\partial }^2r_a}\frac{{\partial }^2E}{\partial r_a\partial r_b}\\ \frac{{\partial }^2E}{\partial r_b\partial r_L}\frac{{\partial }^2E}{\partial r_b\partial r_a}\frac{{\partial }^2E}{{\partial }^2r_b}\end{array}\right).\mathrm{Accordingly},& \left(\mathrm{Eq}.35\right)\\ r_{k+1}=r^{\left(k\right)}-{F\left(r^{\left(k\right)}\right)}^{\left(-1\right)}g^{\left(k\right)}.& \left(\mathrm{Eq}.36\right)\end{array}$

For every r^(k+1), the weighting factor w_j's are updated and the new function E(r) is obtained. The iteration stops when E(r) does not improve.

The algorithms for RBF data fitting may depend on the selection of norms. For color conversions, the norm may be selected considering the visual tolerance of the colors in the CIELab color space and the source of errors in the color conversion process. For example, different norms should be selected for differential color regions, such as

$\begin{array}{cc}\underset{\_}{L_1}\mathrm{norm}\text{:}& \\ {x}_1=\sum _{i=1}^mx_i& \left(\mathrm{Eq}.37\right)\\ L_2\mathrm{norm}\text{:}& \\ {x}_2=\sqrt{\sum _{i=1}^mx_i^2}& \left(\mathrm{Eq}.38\right)\\ L_{\infty }\mathrm{norm}\text{:}& \\ {x}_{\infty }=\underset{1\le x\le m}{\max }x_i.& \left(\mathrm{Eq}.39\right)\end{array}$

The relationship of errors via L₁ norm, L₂ norm, and L_∞ may be stated as: For all e in C[a; b] the inequalities,

∥e∥₁≦(b−a)^1/2∥e∥₂≦(b−a)∥e∥_∞  (Eq. 40)

hold.

This theorem indicates that it small errors are found via L_∞ norm, then the errors via L₁ norm and L₂ norm are also small. Thus, [a, b]=[0, 1], and the CIELab color space may be heuristically chosen in three regions:

• 1. Neutral region: 0≦C*_ab≦7;
• 2. Medium color region: 7<C*_ab≦30; and
• 3. Saturated color region: C*_ab>30,where C*_ab is the chroma of colors in CIELab color space.

For the neutral color region, the visual color difference tolerance is relatively small. Error from the measurements and machine can affect the accuracy of color conversion. The L₁ norm is chosen since it is least sensitive to the outliers. For the medium color range, the color difference tolerance is larger and the errors from the measurements and machine play less of a role affecting the quality of the color conversions. However, control of the color conversion error should be less than a magnitude of 2 so the color difference is less visible. The L₂ norm is chosen to minimize the average error. For the saturated color regions, the errors are generally small compared to the color conversion error of gamut mapping and it is unlikely for the HVS to detect any errors with magnitude less than 2. To minimize the maximum error of the color conversion, the L_∞ norm is generally chosen.

As mentioned, the continuity and smoothness are important to the quality of the color conversion. When adjacent clusters or two adjacent regions have different mapping functions, data in the adjacent clusters may be included in the current cluster. Overlapping ranges specified as δ's for each region are optimized. The δ values are the distances calculated in the CIELab color space. The data regions partitioned for the L₁, L₂, and L_∞ approximation are:

• 1. L₁ approximation: 0≦C*_ab<7+δ_n. Intended range: 0≦C*_ab≦7;
• 2. L₂ approximation: 7−δ_ml<C*_ab<30+δ_mu. Intended range: 7≦C*_ab≦30;
• 3. L_∞ approximation: C*_ab>30−δ_s. Intended range: C*_ab≧30,where δ_n is the extended region for the L₁ norm, δ_ml and δ_mu are the upper and lower extended regions for the L₂ norm, and δ_s is the extended region for the L_∞ norm. For the CMYK to CIELab conversion, the domain space is the undivided CMYK color space. For the Cielab to CMY conversion, the domain space is the CIELab color space divided into three regions based on the selected set of training data described above. Although shown and described with a particular range, the invention is not intended to be so limited. For example, the approximation may be varied as desired to change the granularity of the approximation.

Validation techniques, such as “cross validation” and “bootstraps”, may be used for model accuracy estimation and model selection. Among the cross validation techniques, the “K-fold cross” validation and the “leave-one-out” cross validation may be used. Using the cross validation technique, the absolute accuracy and the bias of variances should be balanced.

In general, several thousand data points are selected for building a color conversion model. A relatively large K number may be chosen to ensure the model accuracy. Different from other cross validation applications, the evaluation of color conversion models involves both numerical analysis and psychophysics analysis. For the numerical analysis, the test data set is collected spanning the input color space. For the perceptual analysis, a number of images may be selected in a visual experiment.

L₁, L₂, and L_∞ approximations may be used to select the K value to the color conversion. For the L₂ approximation, the mean error may be used as the measure for the model selection. And, for the L₁ and L_∞ approximations, the L₁ error and maximum error may be calculated instead, respectively. In any case, the variance as a function of the K value may be obtained to decide the K value. The cross validation technique is described using CMYK to CIELab conversion as an example. First, the modeling set is divided into K-fold partition data sets. For each of K experiments, K−1 folds are used for training with the remaining one for testing. Then, K-fold cross validation is applied to determine the best K value and model. The mean error decreases with the increase of the K value for K<10 and it stays roughly at the same level for K>15. On the other hand, the variance of the mean error stays relatively flat for K<15 and increases with the increase of the K value. Although the maximum error decreases and the mean error stays relatively flat for K≧15, the K value for the least squares approximation for the color conversion is generally 15 due to the increase of the variance.

The CMYK to CIELab conversion may be performed using the RBF data fitting algorithm in the L₂ norm. In this regard, data may be clustered using the combination of the LBG algorithm and the covariance matrix algorithm. Generally, there are two optimization stages. Initially, a relatively good starting function and radius for the R4→R3 optimization is sought. Then, models of R4→R for the CMYK to L*, a*, b* conversions are sought.

In the next modeling stage, f_L, f_a, and f_b are set to multiquadratic with initial r_L, r_a, and r_b. The multidimensional Newton's method is then applied to optimize these parameters. For the CIELab to CMY conversion, both data range (δ_ml and δ_mu) for the least squares approximation and the color conversion model are optimized by letting δ_ml=0, δ_mu=0, and k=0, such that s_k={(La, b)_jk|7≦C*_abjk≦30}, and initiating the clustering algorithm to find the clusters. Then, R3→R is optimized for the CIELab to C, M, and Y conversions, respectively. The multidimensional Newton's method of R3→R3 is performed for the CIELab to CMY conversion and the mean error and maximum error are recorded.

Afterwards, let δ_ml=2.5, δ_mu=2.5, and k=0, such that s_k={(L,a,b)_jk|7−2.5 k≦C*_abjk≦30+2.5 k, k=1,2, . . . , } repeat the optimization process stopping when all data points are included in s_k. The least square errors are obtained by calculating the differences between the original values and the estimated values. The optimized models indicate the Gaussian function for the cyan component, the inverse multiquadratic for the magenta component, and the multiquadratic function for the yellow component, thereby providing inside gamut color conversions from CIELab to CMY, g₀, and CMY to CIELab, f₀.

Overdetermined Linear Systems in the L_∞ Sense

In saturated color areas, the perceptual color difference tolerance is generally higher than the magnitude of machine and instrument errors. Data outliers do not play a significant role in the color conversion. The maximum error is weighted in the color conversion quality assessments. Thus, it is desirable to reduce the maximum approximation error by solving the problem via the L_∞ approximation, also referred as the Chebyshev norm. From norm inequalities if L_∞ norm is controlled, then the L₁ and L₂ error is generally small.

Many algorithms have been developed for this type of application function approximation using the L_∞ norm that are locally equivalent to the exchange algorithm. However, there is limited research on scattered data fitting techniques using the L_∞ norm especially for the large sparse problems. Large sparse problems are often converted to linear programming problems using the linearized sub-problem as the basis. Among these algorithms, the Barrodale and Phillips dual algorithm (the “BP” algorithm) and the Bartels, Conn, and Li primal algorithm (the “BC” algorithm) are suitable algorithms for solving overdetermined linear systems Rn→R in the Chebyshev norm.

The BP and BC algorithms may be suitable in view of the similarity of the formulations with the RBF approach. As the BP and BC algorithms are generally used in Rⁿ→R, the extension of the BP dual algorithm may be extended to R3→R3 for color conversion from CIELab to CMY. In this regard, the Chebyshev norm may be described by letting

Ax=b,   (Eq. 41)

where A=[a₁,a₂, . . . ,a_n] ∈ Rmxm,m>n≧2, and b^T=[β₁, β₂, . . . , β_m] ∈ Rm. Then, the vector x ∈ Rn is determined such that

$\begin{array}{cc}{\mathrm{Ax}-b}_{\infty }=\underset{1\le i\le m}{\max }a_i^Tx-{\beta }_i& \left(\mathrm{Eq}.42\right)\end{array}$

is minimized. L is then defined as a linear space spanned by the functions {φ_i}

L=<φ ₁(z),φ₂(z), . . . ,φ_n(z)>,   (Eq. 43)

where each φ_i is continuous on [a,b],j=1, . . . , n. A function φ ∈ L may then be expressed as

$\begin{array}{cc}\phi \left(z\right)=\sum _{j=1}^n{\alpha }_j{\phi }_j\left(z\right)\in L,& \left(\mathrm{Eq}.44\right)\end{array}$

where z ∈ [a, b]. The RBF is a relatively good example of such a space. Given m>n≧2, data points {(z_i; y_i); i=1; 2; . . . ;m}, α=(α₁, . . . ,α_n) is determined to minimize

$\begin{array}{cc}\begin{array}{c}e\left({\alpha }_1,\dots ,{\alpha }_n\right)={y_i-\sum _{j=1}^n{\alpha }_j{\phi }_j\left(z_i\right)}_{\infty }\\ =\underset{1\le i\le m}{\max }y_i-\sum _{j-1}^n{\alpha }_j{\phi }_j\left(z_i\right).\\ \mathrm{Then},\end{array}& \begin{array}{c}\begin{array}{c}\left(\mathrm{Eq}.45\right)\\ \end{array}\\ \left(\mathrm{Eq}.46\right)\end{array}\\ x=\alpha =\left\{a_1,{\alpha }_2,\dots ,{\alpha }_n\right\}& \left(\mathrm{Eq}.47\right)\\ A=\left(\begin{array}{cccc}{\phi }_1\left(z_1\right)& {\phi }_2\left(z_1\right)& \dots & {\phi }_n\left(z_1\right)\\ \dots & & & \\ \dots & & & \\ {\phi }_1\left(z_m\right)& {\phi }_2\left(z_m\right)& \dots & {\phi }_n\left(z_m\right)\end{array}\right)\mathrm{and}& \left(\mathrm{Eq}.48\right)\\ b={\left(y_1y_2\dots y_m\right)}^T.& \left(\mathrm{Eq}.49\right)\end{array}$

Next, the L_∞ approximation may be converted to a standard form of the linear programming. First, however, the primal form of the problem is established by letting ξ denote the maximum error. For example,

$\begin{array}{cc}\xi =\underset{1\le i\le m}{\max }e_i\left({\alpha }_1,{\alpha }_2,\dots ,{\alpha }_n\right),& \left(\mathrm{Eq}.50\right)\end{array}$

where e is defined in Eq. 41.

The primal problem may then be formulated by minimizing ξ subject to

$\begin{array}{cc}=\xi +\sum _{j-1}^n{\alpha }_j{\phi }_j\left(z_i\right)\ge y_i\mathrm{and}& \left(\mathrm{Eq}.51\right)\\ =\xi -\sum _{j-1}^n{\alpha }_j{\phi }_j\left(z_i\right)\ge -y_i,\mathrm{for}\mathrm{every}i\mathrm{where}& \left(\mathrm{Eq}.52\right)\\ \xi \ge 0\mathrm{and}& \left(\mathrm{Eq}.53\right)\\ {\alpha }_1,{\alpha }_2,\dots ,{\alpha }_n\mathrm{are}\mathrm{unrestricted}.& \left(\mathrm{Eq}.54\right)\end{array}$

The standard form of the corresponding dual linear program may be transformed from the above primal linear program via

$\begin{array}{cc}\max \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i-{\underset{\_}{\tau }}_i\right)y_i\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.55\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i-{\underset{\_}{\tau }}_i\right)\phi j_i=0& \left(\mathrm{Eq}.56\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i+{\underset{\_}{\tau }}_i\right)\le 1& \left(\mathrm{Eq}.57\right)\\ {\underset{\_}{\sigma }}_i,{\underset{\_}{\tau }}_i\ge 0,i=1,2,\dots m,\mathrm{where}{\phi }_{\mathrm{ji}}={\phi }_j\left(z_i\right).& \left(\mathrm{Eq}.58\right)\end{array}$

To understand the BP dual algorithm, an understanding of the exchange algorithm and related theorems which almost all Chebyshev approximation algorithms are based is helpful. In this regard, A may be defined as a linear subspace of C [a,b],f ∈ C [a,b], and L is any closed subset of [a,b], p ∈ A, such that

e*(x)=f(x)−p*(x)   (Eq. 59)

and

L _M ={x ∈ L|e*(x) obtains extreme values}.   (Eq. 60)

p* provides for a minimax solution to be approximation from A to f, if ∀ p ∈ A, such that

e*(x)p(x)=[f(x)−p*(x)]p(x)>0, ∀x ∈ L_M   (Eq. 61)

(e.g., e*(x) and p(x) are of same signs at all extreme points).

In this regard, the “Haar” condition defines a class of functions upon which the Exchange Algorithm can be applied. For example, A may be defined as an (n+1) dimensional linear subspace of C[a, b]. A satisfies the Haar condition if ∀ p ∈ A, the number of roots of the equation

{p(x)=0;a≦x≦b}  (Eq. 62)

is less than the dimension of A.

A derived condition states that if {ξ_i: i=1; 2, . . . , k}⊂[a, b], k≦n then ∃ p ∈ A, such that p changes signs at ξ_i: i=1; 2, . . . , k and p has no other roots. The RBFs in the form of Eq. 11 meet the Harr condition because of the linear part in the equation. Then, let A be an (n+1)-dimensional linear subspace of C[a, b]that satisfies the Haar condition and let f be any function in C[a, b]. Then p* is the minimax approximation from A to f, if there exists (n+2) points {ξ*_i;i=0; 1; . . . ; n+1}, such that the conditions

• 1. a≦ξ*₀<ξ*₁≦ . . . ≦ξ*_n+1≦b,
• 2. |f(ξ*_i)−p*(ξ*_i)|=∥f−p*∥_∞,i=0,1, . . . , n+1 and
• 3f(ξ*_i+1)−p* (ξ*_i+1)=−[f(ξ*_i)−p*(ξ*_i)],i=0, . . . n, alternative signs]] are obtained.

Then, let A be a finite-dimension subspace of C[a, b] that satisfies the Haar condition and let {x_i: i=1; 2; . . . ; m} be a set of distinct points from [a, b], where m is not less than the dimension of A. For any f in C[a, b], let the one-point exchange algorithm be applied to calculate the element of A that minimizes expression

$\begin{array}{cc}\underset{i-1,2,\dots m}{\max }f\left(x_i\right)-p\left(x_i\right),p\in A& \left(\mathrm{Eq}.63\right)\end{array}$

Then, the required approximation to f is obtained in a finite number of iterations.

In color conversion may compose a set of discrete data points. Thus, the Exchange Algorithm may be based on the Minimax Characterization Theorem (MCT) as shown in FIG. 5. For each iteration k, the MCT adjusts a reference {ξ*_i:i=0; 1; . . . ; n+1} and a trial approximation p_k is determined. On a discrete set, f may be known on a set of points {x_i; i=1; 2; : : : ;m}. To determine a trial approximation p_k in each iteration, let

h=f(ξ₀)−p_k(ξ₀)   (Eq. 64)

and solve

f(ξ_i)−p(ξ_i)=(−1)ⁱh,i=0,1,2, . . . ,n+1   (Eq. 65)

The iteration stops when p_k satisfies the conditions in the MCT.

For solving the dual problem, the BP dual algorithm is reviewed. The BP dual algorithm is based on the standard form of the simplex method composed of three stages and introduces artificial variables α_j in Eq. 55 and a slack variable ξ in Eq. 56. The above dual problem thereby becomes

$\begin{array}{cc}\max \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i-{\underset{\_}{\tau }}_i\right)y_i\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.66\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i-{\underset{\_}{\tau }}_i\right){\phi }_{j_i}+{\underset{\_}{\alpha }}_i=0,j=1,2,\dots ,n& \left(\mathrm{Eq}.67\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_i+{\underset{\_}{\tau }}_i\right)+\underset{\_}{\xi }=1& \left(\mathrm{Eq}.68\right)\\ \xi ,{\underset{\_}{\alpha }}_i,-{\underset{\_}{\tau }}_i\ge 0,j=1,2,\dots ,m.& \left(\mathrm{Eq}.69\right)\end{array}$

Denoting the column vectors formed by the coefficients of σ_i, α_i, τ_i and ξ by s_i, t_i, a_j, and w, then

si+t _i=2w  (Eq. 70)

Then, let b denote the right-hand side of Eq. 52 such that

b=w=(0 0 . . . 1)^T=e_n+1^T.   (Eq. 71)

The first two stages of the BP dual algorithm include determining an efficient starting point for the third stage, which is equivalent to the exchange algorithm for the linear minimax approximation. In doing so, the BP dual algorithm uses the table below:

Then, Φ is denoted as an m xn matrix:

$\begin{array}{cc}\Phi =\left(\begin{array}{cc}{\phi }_{1,1}& {\phi }_{2,1}\dots {\phi }_{n,1}\\ {\phi }_{1,2}& {\phi }_{2,2}\dots {\phi }_{n,2}\\ \dots & \\ \dots & \\ {\phi }_{1,m}& {\phi }_{2,m}\dots {\phi }_{n,m}\end{array}\right)\mathrm{and}& \left(\mathrm{Eq}.72\right)\end{array}$

the rank of Φ is denoted as k.

In the first stage, the first k simplex iterations are performed. The algorithm finds the maximum residual at each iteration and reduce it to zero. The nonbasic variables that can enter the basis to become basic variables are σ_j's, and basic variables that can leave the basis to become nonbasic variables are restricted to α_is. Basic variable ξ remains zero in the basis. At the end of stage 1, there are k residues at zero. σ_t that enters the basis corresponds to that with the largest absolute reduced cost (e.g., choosing the largest residual in the primal problem). The pivot may be selected from this column as the largest absolute value corresponding to an α_j in the basis. At the t^th iteration, with the corresponding σ_t entering the basis, there are t zero residuals. For example,

$\begin{array}{cc}{y}_t-\sum _{j=1}^n{\alpha }_j{\phi }_j\left(z_t\right)=\xi =0.& \left(\mathrm{Eq}.73\right)\end{array}$

In the second stage, the (k+1)^th simplex iteration is performed and the variable ξ is forced to leave the basis. This means the error ξ in the primal problem is increased above zero. At the end of the second stage, the k+1 residuals of magnitude equal to resulting value of ξ. The (k+1)^th pivotal column may be chosen also corresponding to the largest absolute marginal cost (e.g., equivalent to choosing the largest residual in the primal problem). The σ_j's and τ_j's are exchanged if the marginal cost is negative. If there are positive values in the first n elements in the pivotal column, these values are changed to negative by adding twice of the corresponding rows to the (n+1)^th pivotal row and changing the sign of those original rows.

The third stage is equivalent to the exchange algorithm for a linear minimax approximation. The pivotal column may be chosen corresponding to the most negative marginal cost or a ratio selection rule. The iteration continues until all marginal cost is nonnegative. Every iteration may increase the value of ξ until ξ=e*, where e* is the minimum value in Eq. 45.

Alternatively, the BC algorithm may be used to determine the starting point for color conversion. In doing so, the primal formulation in Eq. 45 is rewritten by letting

$\begin{array}{cc}{c}_0={\left(\begin{array}{ccccc}1& 0& 0& 0\dots & 0\end{array}\right)}_{n+1}^T& \left(\mathrm{Eq}.74\right)\\ c_i={\left(\begin{array}{c}1\\ -a_i\end{array}\right)}_{n+1}& \left(\mathrm{Eq}.75\right)\\ c_{m+i}={\left(\begin{array}{c}1\\ +a_i\end{array}\right)}_{n+1}& \left(\mathrm{Eq}.76\right)\\ {\delta }_i=-y_i& \left(\mathrm{Eq}.77\right)\\ {\delta }_{m+i}=y_i,i=1,2,\dots ,m,\mathrm{and}& \left(\mathrm{Eq}.78\right)\\ v={\left(\begin{array}{c}\xi \\ \alpha \end{array}\right)}_{n+1}& \left(\mathrm{Eq}.79\right)\end{array}$

The above primal problem is then converted to an unconstrained optimization problem via a piecewise linear penalty function. For example, let μ>0 be a fixed parameter and define

$\begin{array}{cc}p\left(\mathrm{vi},\mu \right)=\mu c_0^Tv-\sum _{j=1}^{2m}\min \left(0,c_j^Tv-{\delta }_j\right)& \left(\mathrm{Eq}.80\right)\end{array}$

For any arbitrary v ∈ Rn+1, Eq. 80 can be expanded into

$\begin{array}{cc}p\left(v,\mu \right)=\mu c_0^Tv-\sum _{j\in I^0}\min \left(0,c_j^Tv-{\delta }_j\right)& \left(\mathrm{Eq}.81\right)\\ -\sum _{j\in I^{+}}\min \left(0,c_j^Tv-{\delta }_j\right)& \left(\mathrm{Eq}.82\right)\\ -\sum _{j\in I^{-}}\min \left(0,c_j^Tv-{\delta }_j\right),\mathrm{where}\text{}I^0=I^0\left(v\right)=\left\{j|c_j^Tv={\delta }_j\right\}=\left\{j_1,j_2,\dots ,j_k\right\},I^{+}=I^{+}\left(v\right)=\left\{j|c_j^Tv>{\delta }_j\right\},\mathrm{and}I^{-}=I^{-}\left(v\right)=\left\{j|c_j^Tv<{\delta }_j\right\}.& \left(\mathrm{Eq}.83\right)\end{array}$

For any d ∈ Rⁿ⁺¹ and any λ≧0 sufficiently small,

$\begin{array}{cc}\begin{array}{c}p\left(v+\lambda d,\mu \right)=p\left(v+\mu \right)+\lambda [\mu c_0^Td-\sum _{j\in I^{-}}^{}c_j^Td-\\ \sum _{j\in I^0}^{}\min \left(0,c_j^Td\right)]\\ =p\left(v+\mu \right)+\lambda h^Td+\lambda \sum _{j\in I^0}^{}{\sigma }_j^{-}c_j^Td,\end{array}& \begin{array}{c}\left(\mathrm{Eq}.84\right)\\ \\ \\ \\ \\ \left(\mathrm{Eq}.85\right)\end{array}\\ \mathrm{where}h=\mu c_0-\sum _{j\in I^{-}}^{}c_j,\mathrm{and}{\sigma }_j^{-}={\sigma }_j^{-}\left(d\right)=\{\begin{array}{cc}0& \mathrm{if}c_j^Td\ge 0\\ 1& \mathrm{if}c_j^Td<0,\end{array}j\in I^0.\mathrm{Then},\mathrm{define}\mathrm{the}\mathrm{matrix}& \\ N=\left[c_{j1},c_{j2},\dots ,c_{\mathrm{jk}}\right]\mathrm{and}\mathrm{let}& \left(\mathrm{Eq}.86\right)\\ P=I-{N\left(N^TN\right)}^{-1}N^T.& \left(\mathrm{Eq}.87\right)\end{array}$

The matrix P is the orthogonal projector onto the null space of N^T. To define search direction d such that p(ν+λd, μ)<p(ν, μ),

d=−Ph   (Eq. 88)

is established. When Ph≠0, apply d defined in Eq. 88 to Eq. 85 such that

$\begin{array}{cc}\begin{array}{c}{c}_j^Td=-c_j^T\mathrm{Ph}\\ =-c_j^T\mathrm{PPh}\\ =-{\left({\mathrm{Pc}}_j\right)}^T\mathrm{Ph}\\ =0\end{array}& \left(\mathrm{Eq}.89\right)\end{array}$

Since P is the orthogonal projector onto the null space of N^T. The Eq. 84 becomes:

p(ν+λd, μ)=p(ν, μ)+λh^T d   (Eq. 90)

h^Td<0, and   (Eq. 91)

p(ν+λd, μ)<p(ν, μ).   (Eq. 92)

Thus, d serves as a descent direction for the function p.

When Ph=0, assuming columns of N are linearly independent, then

$\begin{array}{cc}h=\sum _{i=1}^k{\eta }_ic_{\mathrm{ji}}.\mathrm{and}\mathrm{Eq}.84\mathrm{be}\mathrm{comes}& \left(\mathrm{Eq}.93\right)\\ p\left(v+\lambda d,\mu \right)=p\left(v,\mu \right)+\lambda \sum _{i=1}^k\left[{\eta }_i+{\sigma }_{\mathrm{ji}}^{-}\right]c_j^Td.& \left(\mathrm{Eq}.94\right)\end{array}$

In the sub case of a: η_i*<0, for some i* ∈ I⁰, d is chosen such that

c _j ^T d=0,j ∈ I⁰,j≠i*   (E1. 95)

c_i^T*d=1   (Eq. 96)

This implies that σ_j⁻=0<0, j ∈ I⁰. Because

h ^T d=η _i *c _i ^T *d=η<0,   (Eq. 97)

then

p(ν+λd, μ)<p(ν, μ).   (Eq. 98)

In the sub case of b: η_i*<0 for all i* ∈ I⁰, if no constraints are violated, it is the optimal. Otherwise, reduce μ and start another iteration. For example,

{choose step size λ}

$\begin{array}{cc}{\Lambda }^{\left(1\right)}\stackrel{\mathrm{def}}{=}\left\{\lambda \lambda =\frac{{\delta }_j-c_j^Tv}{c_j^Td},\lambda >0,j\in I^{+}\bigcup I^{-}\right\},& \left(\mathrm{Eq}.99\right)\\ \lambda =\min {\Lambda }^{\left(1\right)}& \left(\mathrm{Eq}.100\right)\end{array}$

and update ν=ν+λd.

As an example, the L_∞ norm approximation may be applied to color conversions from CMYK to CIELab and CIELab to CMY. The formulation of this color conversion in the primal form is described in the following using the CIELab to CMY color conversion by minimizing

$\begin{array}{cc}\xi ={\xi }_c+{\xi }_m+{\xi }_y\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.101\right)\\ {\xi }_c+\sum _{j=1}^{n+4}{\alpha }_{\mathrm{cj}}F_{\mathrm{cj}}\left(z_i\right)\ge c_i& \left(\mathrm{Eq}.102\right)\\ {\xi }_c-\sum _{j=1}^{n+4}{\alpha }_{\mathrm{cj}}F_{\mathrm{cj}}\left(z_i\right)\ge c_i& \left(\mathrm{Eq}.103\right)\\ {\xi }_m+\sum _{j=1}^{n+4}{\alpha }_{\mathrm{mj}}F_{\mathrm{mj}}\left(z_i\right)\ge m_i& \left(\mathrm{Eq}.104\right)\\ {\xi }_m-\sum _{j=1}^{n+4}{\alpha }_{\mathrm{mj}}F_{\mathrm{mj}}\left(z_i\right)\ge m_i& \left(\mathrm{Eq}.105\right)\\ {\xi }_y+\sum _{j=1}^{n+4}{\alpha }_{\mathrm{yj}}F_{\mathrm{yj}}\left(z_i\right)\ge y_i& \left(\mathrm{Eq}.106\right)\\ {\xi }_y-\sum _{j=1}^{n+4}{\alpha }_{\mathrm{yj}}F_{\mathrm{yj}}\left(z_i\right)\ge y_i,& \left(\mathrm{Eq}.107\right)\end{array}$

where

ε≧0,

• α₁, α₂, . . . α_n+4 are unrestricted,

z _i ={L, a, b}i, i=1, . . . , m,

F _ji=[1, z_i, φ₁(z_i), φ₂(z_i), . . . , φ_n(z_i)]_4+n,

φ_ji=φ_j(z_i),

φ_j is one of the RBFs, n is the number of clusters, and m is the number of color data points.

The problem may then be rewritten by minimizing c_o^Tν subject to

$\begin{array}{cc}c_j^Tv\ge {\delta }_j,j=1,2,\dots 6m\mathrm{where}c_0={\left(\begin{array}{cccccccccc}1& 0& \dots & 1& 0& \dots & 1& 0& \dots & 0\end{array}\right)}_{3n+15}^Tv={\left(\begin{array}{cccccccccc}{\xi }_c& {\alpha }_{c1}& \dots & {\alpha }_{\mathrm{cn}}& {\xi }_m& {\alpha }_{m1}& \dots & {\alpha }_{\mathrm{mn}}& {\xi }_y& {\alpha }_{y1}\\ \dots & {\alpha }_{\mathrm{yn}}& & & & & & & & \end{array}\right)}_{3n+15}^Tc_{i}={\left(\begin{array}{cccccc}1& -{\alpha }_{\mathrm{ci}}& \dots & 0& \dots & 0_{2n+8}\end{array}\right)}_{3n+15}^Tc_{m+i}={\left(\begin{array}{cccccc}1& {\alpha }_{\mathrm{ci}}& \dots & 0& \dots & 0_{2n+8}\end{array}\right)}_{3n+15}^Tc_{2m+i}={\left(\begin{array}{cccccccccc}0& \dots & 0_{n+4}& \dots & 1& \dots & -{\alpha }_{mi}& 0& \dots & 0_{n+4}\end{array}\right)}_{3n+15}^Tc_{3m+i}={\left(\begin{array}{ccccccccc}0& \dots & 0_{n+4}& \dots & 1& {\alpha }_{mi}& 0& \dots & 0_{n+4}\end{array}\right)}_{3n+15}^Tc_{4m+i}={\left(\begin{array}{ccccc}0& \dots & 0_{2n+8}& 1& -{\alpha }_{\mathrm{yi}}\end{array}\right)}_{3n+15}^Tc_{5m+i}={\left(\begin{array}{ccccc}0& \dots & 0_{2n+8}& 1& {\alpha }_{\mathrm{yi}}\end{array}\right)}_{3n+15}^T{\delta }_i=-C_i{\delta }_{m+i}=C_i{\delta }_{2m+i}=-M_i{\delta }_{3m+i}=M_i{\delta }_{4m+i}=-Y_i{\delta }_{5m+i}=Y_i,\mathrm{and}i=1,\dots ,m.& \left(\mathrm{Eq}.108\right)\end{array}$

The primal problem may then be divided into steps using the modified BC algorithm. For example, the primal problem with constraints may be converted to an unconstrained problem via a piecewise linear penalty function by letting μ>0 be a fixed parameter. Next,

$\begin{array}{cc}p\left(v,\mu \right)=\mu c_0^Tv-\sum _{j=1}^{6m}\min \left(0,c_j^Tv-{\delta }_j\right)\mathrm{is}\mathrm{defined}.\mathrm{For}\mathrm{any}\mathrm{arbitrary}v\in {3n+12}^{},& \left(\mathrm{Eq}.109\right)\\ p\left(v,\mu \right)=\mu c_0^Tv-\sum _{j\in I^0}\min \left(0,c_j^Tv-{\delta }_j\right)-\sum _{j\in I^{+}}\min \left(0,c_j^Tv-{\delta }_j\right)-\sum _{j\in I^{-}}\min \left(0,c_j^Tv-{\delta }_j\right),\mathrm{where}I^0=I^0\left(v\right)=\left\{jc_j^Tv={\delta }_j\right\}=\left\{j_1,j_2,\dots ,j_k\right\},I^{+}=I^{+}\left(v\right)=\left\{jc_j^Tv>{\delta }_j\right\},\mathrm{and}I^{-}=I^{-}\left(v\right)=\left\{jc_j^Tv<{\delta }_j\right\}.& \left(\mathrm{Eq}.110\right)\end{array}$

Then, the sets I⁰, I⁺, and I⁻ are determined and the search direction and step size are updated.

As an example of color conversion in dual for CIELab to CMY,

$\begin{array}{cc}\mathrm{Maximize}\sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{ci}}-{\underset{\_}{\tau }}_{\mathrm{ci}}\right)c_i+\left({\underset{\_}{\sigma }}_{\mathrm{mi}}-{\underset{\_}{\tau }}_{mi}\right)m_i+\left({\underset{\_}{\sigma }}_{\mathrm{yi}}-{\underset{\_}{\tau }}_{\mathrm{yi}}\right)y_i\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.111\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{ci}}+{\underset{\_}{\tau }}_{\mathrm{ci}}\right)F_{C_{\mathrm{ji}}}^T=0& \left(\mathrm{Eq}.112\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{ci}}-{\underset{\_}{\tau }}_{\mathrm{ci}}\right)\le 1& \left(\mathrm{Eq}.113\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{Mi}}+{\underset{\_}{\tau }}_{\mathrm{Mi}}\right)F_{M_{\mathrm{ji}}}^T=0& \left(\mathrm{Eq}.114\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{Mi}}-{\underset{\_}{\tau }}_{\mathrm{Mi}}\right)\le 1& \left(\mathrm{Eq}.115\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{Yi}}-{\underset{\_}{\tau }}_{\mathrm{Yi}}\right)F_{Y_{\mathrm{ji}}}^T=0& \left(\mathrm{Eq}.116\right)\\ \sum _{i=1}^m\left({\underset{\_}{\sigma }}_{\mathrm{Yi}}+{\underset{\_}{\tau }}_{\mathrm{Yi}}\right)\le 1\mathrm{and}& \left(\mathrm{Eq}.117\right)\\ {\underset{\_}{\sigma }}_i,{\underset{\_}{\tau }}_i\ge 0,i=1,2,\dots ,m,\mathrm{where}z_i={\left\{L,a,b\right\}}_i,i=1,2,\dots ,mF_{\mathrm{ji}}={\left[1,z_i,{\varphi }_1\left(z_i\right),{\varphi }_2\left(z_i\right),\dots ,{\varphi }_n\left(z_i\right)\right]}_{n+4}{\varphi }_{\mathrm{ji}}={\varphi }_j\left(z_i\right),& \left(\mathrm{Eq}.118\right)\end{array}$

φ_j is one of the RBFs, n is the number of clusters, and m is the number of color data points.

The CIELab to CMY conversion may then be represented by the initial simplex table:

and the rank of the matrix F_n×m, k is thereby determined.

Thereafter, the first step BP simplex algorithm may be performed via k simplex iterations for each of F_Cn×m for the cyan component, F_Mn×m for the magenta component, and F_Yn×m for the yellow component. The vectors s_C are allowed to enter the basis and the vectors a_C can leave the basis for F_Cn×m. The vectors s_M are allowed to enter the basis and the vectors a_M can leave the basis for F_Mn×m. The vectors s_Y are allowed to enter the basis and the vectors a_Y can leave the basis for Y_n×m.

The second step of the BP simplex algorithm may be performed by forcing ξ_C, ξ_M, and ξ_Y to leave the basis. The corresponding vectors entering the basis are sC, s_M, and s_Y respectively. If the corresponding marginal cost of s_C, s_M, or s_Y is positive, then si is changed to t_i using

t _i=2w3_n−4−s_i,

where

w _C=(0 0 . . . 0 1 0 0)^T=e_3n+1^T,

w _M=(0 0 . . . 0 0 1 0 0)^T=e_3n+2^T, and

w _Y=(0 . . . 0 0 0 1 0)^T=e_3n+3^T.

If s_i is the pivotal column, r is the row vector of pivotal row. For any j^th row vector r_j containing s_ij>0, where j≠3n+1 for F_Cn×m, j≠3n+2 for F_Mn×m, and j≠3n+3 for F_Yn×m, change r to r+2r_j and change r_j to −r_j. The third step of the BP simplex iterations may then be performed until all the marginal costs are non-negative.

Color Conversion Using the L₁ Norm Approximation

Color conversion in neutral color areas plays an important role in color reproduction quality. The human visual system (HVS) is sensitive to color changes in neutral color areas. Additionally, the color conversion errors due to machine stability, paper uniformity, and instrument accuracy are more profound relative to the human color difference tolerances. Thus, there are more chances for the existences of outliers using color conversion. A color conversion should minimize the negative impact of outliers. The L₁ norm approximation may assist in the region due to the robustness to outliers.

Among all the L₁ algorithms, the Barrodale and Roberts' primal algorithm (BR algorithm) is relatively robust. Similar to the problem setup for the L_∞, the overdetermined linear system of the general L₁ approximation problem may be stated as follows:

Ax=b,   (Eq. 119)

where

A=[a ₁ , a ₂ , . . . , a _n] ∈ ^m×n, (m>n>2)

and

b ^T=[β₁, β₂, . . . , β_m]∈ ^m.

The L₁ norm approximation is used to determine vector x ∈ ⁿ such that

$\begin{array}{cc}{\mathrm{Ax}-b}_1=\sum _{i=1}^m\sum _{j=1}^na_j^Tx_i-{\beta }_i& \left(\mathrm{Eq}.120\right)\end{array}$

is minimized. In this regard, L is a linear space spanned by function φ_i

L=<φ ₁(z), φ₂(z), . . . , φ_n(z)>,   (Eq. 121)

where each φ_i is continuous on [a, b], i=1, . . . , n. The function φ may be expressed as

$\begin{array}{cc}\phi \left(z\right)=\sum _{i=1}^n{\alpha }_i{\phi }_i\left(z\right)\in L,& \left(\mathrm{Eq}.122\right)\end{array}$

where z ∈ [a,b] ⊂ ^k.

Given m>n≧2, data points (z_i, y_i), a=(a₁, . . . a_n) is determined to minimize

$\begin{array}{cc}e\left({\alpha }_1\dots {\alpha }_n\right)={y_i-\sum _{j=1}^n{\alpha }_j{\phi }_j\left(z_i\right)}_1& \left(\mathrm{Eq}.123\right)\\ =\sum _{j=1}^my_j-\sum _{j=1}^n{\alpha }_j{\phi }_j\left(z_i\right)& \left(\mathrm{Eq}.124\right)\end{array}$

The equation then becomes:

$\begin{array}{cc}x=aA=\left(\begin{array}{cccc}{\phi }_1\left(z_1\right)& {\phi }_2\left(z_1\right)& \dots & {\phi }_n\left(z_1\right)\\ \dots & & & \\ \dots & & & \\ {\phi }_1\left(z_m\right)& {\phi }_2\left(z_m\right)& \dots & {\phi }_n\left(z_m\right)\end{array}\right)b={\left(y_1y_2\dots y_m\right)}^T& \left(\mathrm{Eq}.125\right)\end{array}$

In the RBF form, the function becomes

$\begin{array}{cc}\phi \left(z\right)={\alpha }_0+\sum _{p=1}^k{\alpha }_pz\left(p\right)-\sum _{j=n+k}^{n+k}{\alpha }_j\phi \left(z-c_j\right),\mathrm{where}x=\left({\alpha }_0,{\alpha }_1,\dots ,{\alpha }_k,{\alpha }_{k+1},\dots ,{\alpha }_{n+k}\right)\mathrm{and}A=(\begin{array}{ccccccccc}1& z\left(1\right)& z_1\left(2\right)& \dots & z_1\left(k\right)& {\phi }_1\left(z_1\right)& {\phi }_2\left(z_1\right)& \dots & {\phi }_n\left(z_1\right)\\ \dots & & & & & & & & \\ \dots & & & & & & & & \\ \dots & & & & & & & & \\ 1& z_m\left(1\right)& z_m\left(2\right)& \dots & z_m\left(k\right)& {\phi }_1\left(z_m\right)& {\phi }_2\left(z_m\right)& \dots & {\phi }_n\left(z_m\right)\end{array}).& \left(\mathrm{Eq}.126\right)\end{array}$

The L₁ and L_∞ norms in C[a, b] and in R_n are not strictly convex. Accordingly, A is a linear subspace of either the L₁ or the L_∞ normed linear space, as the uniqueness of best approximations from A to f depends on properties of A and f. p* in A may be the better L₁ approximation function to f. A sign function is thus defined as:

$\begin{array}{cc}{s}^{*}\left(x\right)=\{\begin{array}{cc}-1,& f\left(x\right)<p^{*}\left(x\right)\\ 0,& f\left(x\right)=p^{*}\left(x\right)\\ 1,& f\left(x\right)>p^{*}\left(x\right)\end{array}& \left(\mathrm{Eq}.127\right)\end{array}$

A sufficient condition for the function p* is the L₁ approximation from A to f which may be defined by letting A be a linear subspace of C[a, b], f be any function in C[a, b], and p* be any element of A, such that the set

L={x:f(x)=p*(x), a≦x≦b}  (Eq. 128)

is either empty or is composed of a finite number of intervals and discrete points. Then p* provides an L₁ norm approximation form A to f, if the inequality

|∫_a^bs*(x)p(x)dx}|≦∫_L|p(x)|dx (4.7)   (Eq. 129)

is satisfied for all p in A, where s* is the function of Eq. 127.

For discrete L₁ approximation, let the function values {f(x_t); t=1, 2, . . . , m} and the fixed positive weights be {w_t; t=1, 2, . . . , m}. A is a linear space of functions that may be defined on the point set {x_t; t=1, 2, . . . , m}, and p* is any element of A. Then, ε contains the points of {x_t; t=1, 2, . . . , m} that satisfy the condition

p*(x_t)=f(x_t)   (Eq. 130)

and s* is the sign function

$\begin{array}{cc}{s}^{*}\left(x\right)=\{\begin{array}{cc}-1,& f\left(x_t\right)<p^{*}\left(x_t\right)\\ 0,& f\left(x_t\right)=p^{*}\left(x_t\right)\\ 1,& f\left(x_t\right)>p^{*}\left(x_t\right)\end{array}& \left(\mathrm{Eq}.131\right)\end{array}$

t=1, 2, . . . , m. p* becomes the function in A that minimizes the expression

$\begin{array}{cc}\sum _{t=1}^mw_tf\left(x_t\right)-p\left(x_t\right),p\in A& \left(\mathrm{Eq}.132\right)\end{array}$

if the inequality

$\begin{array}{cc}\sum _{t=1}^mw_ts^{*}\left(x_t\right)p\left(x_t\right)\le \sum _{x_t\in L}^mw_tp\left(x_t\right)& \left(\mathrm{Eq}.133\right)\end{array}$

holds for all p in A.

In order to improve the discrete L₁ approximation p*, the inequality in Eq. 133 is tested to be satisfied for every element p in A. An optimal function p* can be obtained by searching for suitable interpolation points in L. For example, let the function values {f(x_t); t=1, 2, . . . , m} and the fixed positive weights be {w_t; t=1, 2, . . . , m}. Then, A is a linear subspace of R^m, where the component of each vector p in A have the values {p(x_t); t=1, 2, . . . , m}. An element p* in A minimizes Eq. 132, and the property that the zero vector is the element p in A that satisfies the conditions {p(x_t)=0; x_t*L}, where the set L is defined in Eq. 128

With this is mind, the discrete L₁ approximation calculation becomes a linear programming problem. For example with φ={φ_i; i=0, . . . , n} as a basis of the space A of approximation, the Eq. 133 becomes

$\begin{array}{cc}\sum _{t=1}^mw_ty_t-\sum _{i=0}^n{\lambda }_i{\phi }_i\left(x_t\right)& \left(\mathrm{Eq}.134\right)\end{array}$

where y_t is the output value corresponding to each x_t. Define the bounds {u₁≧0} and {v_t≧0} may be defined such that

$\begin{array}{cc}-v_t\le y_t-\sum _{i=0}^n{\lambda }_i{\phi }_i\left(x_t\right)\le u_t& \left(\mathrm{Eq}.135\right)\end{array}$

for {t=1, 2, . . . , m}. Then the linear programming problem is

$\begin{array}{cc}\mathrm{Minimize}\sum _{t=1}^mw_t\left(u_t+v_t\right)\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.136\right)\\ -v_t\le y_t-\sum _{i=0}^n{\lambda }_i{\phi }_i\left(x_t\right)\le u_t,t=1,2,\dots m\mathrm{ut}\ge 0\mathrm{vt}\ge 0\mathrm{for}t=1,2,\dots ,m.& \left(\mathrm{Eq}.137\right)\end{array}$

The algorithm may be described as:

$\begin{array}{cc}{f}_i-\sum _{j=1}^n{\alpha }_j{\phi }_{j,i}=u_i-v_i,& \left(\mathrm{Eq}.138\right)\end{array}$

where {u_i, v_i: i=1, 2, . . . , m} are nonnegative variables. The nonnegative variables {b_j, c_j: j 32 1, 2, . . . , n} and the weights a_j=b_j−c_j may then be defined. The primal problem becomes

$\begin{array}{cc}\mathrm{Minimizing}\sum _{i=1}^m\left(u_i+v_i\right)\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.139\right)\\ f_i=\sum _{j=1}^n\left(b_j-c_j\right){\phi }_{j,i}+u_i-v_i,i=1,2,\dots ,m\mathrm{and}& \left(\mathrm{Eq}.140\right)\\ b_j,c_j,u_i,v_i\ge 0.& \left(\mathrm{Eq}.141\right)\end{array}$

The algorithm may be implemented in two stages: For example, for the first n iterations, the pivotal columns are restricted to b_j and c_j. The vector entering the basis is the one with the largest nonnegative marginal cost, (e.g., Σ_i=1^mφ_j,i). The vector leaving the basis is chosen among v_i and u_i for the one causing the maximum reduction in the objective function. Next, the nonbasic u_i or v_i are interchanged with the basic u_i or v_i. Neither b_j or c_j is allowed to leave the basis. The criteria for selecting vectors leaving and entering the basis are the same as that in the first stage. Otherwise, the basic vectors b_j or c_j are interchanged with the corresponding nonbasic vectors c_j or b_j. On making this algorithm efficient, the vector chosen to leave the basis is the one that causes the maximum reduction in the objective function.

On the Multidimensional Neutral Color Conversion Problem, the BR algorithm may be applied to the neutral color conversion by letting z_i={(L, a, b)_i i=1, 2, . . . , m}. As an example, the color conversion from CIELab to CMY is formulated:

$\begin{array}{cc}\mathrm{Minimizing}\xi =\sum _{i=1}^mu_{\mathrm{ci}}+v_{\mathrm{ci}}+u_{\mathrm{mi}}+v_{mi}+u_{\mathrm{yi}}+v_{\mathrm{yi}}\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.142\right)\\ \sum _{j=1}^{n+4}\left(a_{\mathrm{cj}}-b_{\mathrm{cj}}\right)F_{\mathrm{cj}}\left(z_i\right)+u_{\mathrm{ci}}+v_{\mathrm{ci}}=C_i& \left(\mathrm{Eq}.143\right)\\ \sum _{j=1}^{n+4}\left(a_{\mathrm{mj}}-b_{\mathrm{mj}}\right)F_{\mathrm{mj}}\left(z_i\right)+u_{\mathrm{mi}}+v_{\mathrm{mi}}=M_i& \left(\mathrm{Eq}.144\right)\\ \sum _{j=1}^{n+4}\left(a_{\mathrm{yj}}-b_{\mathrm{yj}}\right)F_{\mathrm{yj}}\left(z_i\right)+u_{\mathrm{yi}}+v_{\mathrm{yi}}=Y_i\mathrm{where}a_{\mathrm{cj}},b_{\mathrm{cj}},a_{\mathrm{mj}},b_{\mathrm{mj}},a_{\mathrm{yj}},b_{\mathrm{yj}},u_i,v_i\ge 0\mathrm{and}F_{\mathrm{ji}}={\left[1,z\left(i\right),{\varphi }_1\left(z_i\right),{\varphi }_2\left(z_i\right),\dots ,{\varphi }_n\left(z_i\right)\right]}_{n+4}& \left(\mathrm{Eq}.145\right)\end{array}$

for the C, M, and Y components respectively.

When performing the two-stage BR algorithm, the vectors relating to C, M, and Y are generally be exchanged with the vectors of C, M, and Y respectively. This yields an exemplary condensed initial simplex table for CIELab to CMY conversion as shown below.

One Dimensional Color Manifolds in the CMYK Color Space

One of the more difficult and challenging tasks in color printing is reducing the cost of printing while maintaining quality. Traditionally, the printer color conversion from CMY (or CIELab) to CMYK has been achieved (UCR) and (UCA). As discussed, UCR is the process of removing an equal amount of cyan, magenta, and yellow, and adding the same amount of black. More explicitly, the color components are defined as:

CMY={C,M, Y},CMYK={C,M, Y,K}, and Lab={L, a, b}.

The smallest value of the C, M, Y components is formed as K=mint{C, M, Y} and K is used to convert a CMY value to a CMYK value as follows:

C′=C−K, M′=M−K, and Y′=Y−K.

Because of the nonlinearity of toners/inks, this UCR model is in accurate, and introduces unacceptable color conversion errors. Colors resulting from the UCR process are generally dull, muddy, and hue shifted. To mitigate this problem, the UCR process is followed by the UCA correction which consists of adding a relatively small amount of cyan, magenta, and yellow back to the C′, M′, Y′ values respectively to make colors richer and less hue shifted. However, this method is empirical and labor intensive and the conversion error is still quite large.

A solution to this color toner/ink reduction problem consists of approximating a nonlinear continuous and differentiable functions mapping from CIELab to CMY and CMYK to CIELab. To optimize CMY to CMYK conversion, it is assumed that accurate continuous functions describing CMYK to CIELab conversion and CIELab to CMY conversion are as follows:

g(L,a,b)→(C,M,Y)∈(⁺, ⁺, ⁺) and f(C,M, Y)→(L,a,b) .

For a CIELab value within the printer gamut, Lab₀=(L₀, a₀, b₀), g(L₀, a₀, b₀)=(C₀,M₀, Y₀).If C₀>0 & M₀>0 & Y₀>0, a (C₀′,M₀′,Y₀′, K₀′) value may be determined such that

$\underset{\underset{X}{}}{f\left(C_0^{\prime },M_0^{\prime },Y_0^{\prime },K_0^{\prime }\right)}\approx \underset{\underset{Y}{}}{\left(L_0,a_0,b_0\right)}.$

Then, the CMYK value x_n=(C_n′,M_n′,Y_n′,K_n′) may be approximated iteratively such that, for n large enough,

∥f(x_n)−y_n∥≦ε  (Eq. 146)

for some prescribed tolerance ε. Newton's method may be used if an initial point is close to the minimizer algorithm. To determine an initial point for Newton's method a value x₀ is determine such that

err=∥f(x₀)−y₀∥≦tol,   (Eq. 147)

where tol is an initial acceptable tolerance.

One of the problems in traditional GCR algorithms is that the CIELab value having equal amounts of cyan, magenta, and yellowtoner/ink is often not equal to the CIELab value of that same amount of black toner/ink. In other words, the CIELab value of K generally corresponds to unequal amounts of C, M, Y. If the equal amount of C, M, Y, K values are used as the initial point to Newton's algorithm, the above inequality may not be satisfied. A better initial point may be obtained by removing unequal amounts of cyan, magenta, and yellow from the CMY value respectively, and adding the K value. The maximum error using this process may be improved from the typical UCR error value of 14.42 to 10.50.

On this region, for any K, the associated CIELab value is computed using

f(0, 0, 0,K)→(L^K, a^K, b^K)   (Eq. 148)

Then, the CMY value corresponding to this K is determined using

g(L^K, a^K, b^K)→(C^K, M^K, Y^K)   (Eq. 149)

to establish an equivalence

K(C^K,M^K, Y^K).   (Eq. 150)

This repeated for a sequence of K values span the range of (0, 1). For any given (C,M, Y) value, there exists a maximum K_p value whose corresponding (C^Kp,M^Kp, Y^Kp) satisfies

C=C^Kp or M=M^Kp or Y=Y^Kp   (Eq. 151)

A sequence of K values is determined such that

K={K _i ≦K _p , i=1, 2, . . . , p}.   (Eq. 152)

The corresponding (C^Ki,M^Ki, Y^Ki)'s can also be removed from the (C, M, Y) value. Then, the CMY value (C, M, Y) is converted to a sequence of CMYK values such that

CMYK′={(C^Ki,M^Ki,Y^Ki), i=1,2, . . . , p}.   (Eq. 153)

where C^K′i=C−C_K′i, M^K′i=M−M_K′i, and Y^K′i=Y−Y^K′_i, which yields an optimizing GCR algorithm.

Given an initial CMY, the optimized GCR algorithm may determine an associated (C′,M′,Y′, K′) such that the quality

∥f(C′,M′,Y′,K′)−f(C,M,Y,0)∥  (Eq. 154)

is relatively small. With an initial point close enough to the minimizer, the multidimensional Newton's method may be used to determine the optimal value of (C′,M′,Y′,K′). For example, let

$\begin{array}{cc}f=\left(\begin{array}{c}{f}_L\\ f_a\\ f_b\end{array}\right),x=\left(\begin{array}{c}C\\ M\\ Y\\ K\end{array}\right),y=\left(\begin{array}{c}L\\ a\\ b\end{array}\right),y=\left(\begin{array}{c}{L}_0\\ a_0\\ b_0\end{array}\right),& \left(\mathrm{Eq}.155\right)\end{array}$

where y₀ is the corresponding CIELab value of the original CMY color (e.g., the desired CIELab value for the optimizer of the (C′,M′,Y′,K′)). f(x) then represents the mapping from CMYK to CIELab and expresses as a RBF. For example,

$\begin{array}{cc}{f}_j\left(x\right)=A_jx+a_{j0}+\sum _{i=1}^Nw_{\mathrm{ji}}{\phi }_j\left(x-c_i\right),& \left(\mathrm{Eq}.156\right)\end{array}$

where j=1, 2, 3 correspond to {f_Lf_a,f_b}. Then, f(x+h)≈f(x)+f′(x)h, where f′(x) is in the form of Jacobian matrix

$\begin{array}{cc}J=f^{\prime }\left(x\right)=\left(\begin{array}{cccc}\frac{\partial f_L}{\partial C}& \frac{\partial f_L}{\partial M}& \frac{\partial f_L}{\partial Y}& \frac{\partial f_L}{\partial K}\\ \frac{\partial f_a}{\partial C}& \frac{\partial f_a}{\partial M}& \frac{\partial f_a}{\partial Y}& \frac{\partial f_a}{\partial K}\\ \frac{\partial f_b}{\partial C}& \frac{\partial f_b}{\partial M}& \frac{\partial f_b}{\partial Y}& \frac{\partial f_b}{\partial K}\end{array}\right).& \left(\mathrm{Eq}.157\right)\end{array}$

For each iteration k, x_k+1=x_k−J(x_k)⁻¹(f(y_k)−f(y₀)). This iteration stops when ∥(f(x_n)−y₀)∥≦10⁻¹².

The dimension of the Jacobian matrix is generally 3×4. In order to change it to an invertible Jacobian matrix. The black value of the (C,M, Y,K) color remains unchanged. The Jacobian matrix in Equation 149 is then changed to

$\begin{array}{cc}J=f^{\prime }\left(x\right)=\left(\begin{array}{ccc}\frac{\partial f_L}{\partial C}& \frac{\partial f_L}{\partial M}& \frac{\partial f_L}{\partial Y}\\ \frac{\partial f_a}{\partial C}& \frac{\partial f_a}{\partial M}& \frac{\partial f_a}{\partial Y}\\ \frac{\partial f_b}{\partial C}& \frac{\partial f_b}{\partial M}& \frac{\partial f_b}{\partial Y}\end{array}\right).& \left(\mathrm{Eq}.158\right)\end{array}$

Solving the Equation 149 using SVD obtains

$\begin{array}{cc}{x}_{k+1}=\left(\begin{array}{c}{C}_{k+1}\\ M_{k+1}\\ Y_{k+1}\\ K\end{array}\right).& \left(\mathrm{Eq}.159\right)\end{array}$

At each iteration, check if the C, M, Y values in the converted (C, M, Y K) analyzer to determine whether they are in the range of [0, 1], (e.g., if x_k+1ⁱ≧1 or x_k+1ⁱ≦0, where i=1, 2, 3 correspond to C, M and Y component respectively). If the value of the i^th component is out of the range,

$\begin{array}{cc}{x}_{k+1}^i=\{\begin{array}{c}1,x\left(i\right)\ge 1\\ 0,x\left(i\right)\le 0.\end{array}& \left(\mathrm{Eq}.160\right)\end{array}$

The i^th component remains unchanged for the rest of the iterations. The column that corresponds to this component from the Jacobian matrix J is then removed, and the constraint on the black component is relaxed by adding the partial derivatives with respect to black in the matrix J. For example, if x(2)≧1, the magenta component is removed. The Jacobian matrix J_k−1 becomes

$\begin{array}{cc}{J}_{k+1}=\left(\begin{array}{ccc}\frac{\partial f_L}{\partial C}& \frac{\partial f_L}{\partial M}& \frac{\partial f_L}{\partial Y}\\ \frac{\partial f_a}{\partial C}& \frac{\partial f_a}{\partial M}& \frac{\partial f_a}{\partial Y}\\ \frac{\partial f_b}{\partial C}& \frac{\partial f_b}{\partial M}& \frac{\partial f_b}{\partial Y}\end{array}\right)\mathrm{and}& \left(\mathrm{Eq}.161\right)\\ x_{k+1}=\left(\begin{array}{c}{C}_{k+1}\\ 1.0\\ Y_{k+1}\\ K_{k+1}\end{array}\right).& \left(\mathrm{Eq}.162\right)\end{array}$

Generally, any given CMY value of (C>0 & M>0 & Y>0) converges to a CMYK value with the CIELab value difference practically zero.

As described above, each black level K, (e.g., CMYK=(0, 0, 0, K) can be mapped to a CMY value of (C_k,M_k, Y_k). For any CMY value (C>0 & M>0 & y>0), there exists a set of values {K_i, i=1, 2, . . . , p} in which a corresponding {C^ki,M^ki, Y^ki, i=1, 2, . . . , p} may be removed from C,M, and Y, (e.g.,C≧C^ki & M≧M^ki & Y≧Y^ki, i=1, 2, . . . , p). Generally, for one CMY value, there exists a set of CMYK values {C−C^ki,M−M^ki, Y−Y^ki,K_i, i=1, 2, . . . , p} that can be used for the initial points to the multidimensional Newton's optimization algorithm. Thereafter, the set of values f⁻¹(L₀, a₀, b₀) may be determined as a one-dimensional manifold based on the pre-image theorem.

For example, X→Y be a smooth map between manifolds. Thus, a point is a regular value of f if for all x: x ∈ f⁻¹ (y) and the map df: T_xX→T_yY is subjective. Here, T_xX and T_yY are the tangent spaces of X and Y at the points x and y. Again, with f: X→Y being a smooth map, y ∈ Y may be a regular value of f. Accordingly, x: x ∈ f⁻¹ (y) is a submanifold of X with dim f⁻¹ (y)=dimX−dimY.

For color conversion, A={set of elements of CMY K} and B={set of elements of CIELab}. A function f=(f_L,f_a,f_b) that maps A to B is in the form of RBFs described in Eq. 156 and f is continuous and differentiable. For a CIELab value y ∈ B, if the Jacobian matrix described in Eq. 157 at y is full rank, (e.g., rank is equal to 3), then y is a regular point of f. Generally, there exists a manifold S=f⁻¹(y), S ⊂A with the dimension of one.

These one-dimensional manifolds based on the pre-image theorem are then determined. For example, for any CIELab value y that is inside the printer gamut, obtain the CMY value using g(y)→(C, M, Y). Then, determine whether min(C, M, Y)=0. If so y is not a regular point since its Jacobian matrix will be rank deficient. Otherwise, if min(C, M, Y)>0, the initializing algorithm is performed to obtain the sequence of the CMYK value CMYK′ described in Eq. 155.

The Jacobian matrix is monitored at each starting point to decide if y is a regular point. If y is a regular point, the optimized GCR algorithm is performed to obtain optimized value for each starting point, (e.g., obtain the 1-dimensional manifold S⊂A, and S=f⁻¹(y)). When there exists a level set of CMYK values S=f⁻¹(y) that correspond to the same CIELab value, y ∈ B, the optimal solution for the toner/ink selection can be obtained for a given objective function.

The objective function can be toner/ink cost based. That is, the objective function is to minimize the cost of printing or toner/ink coverage. For example, min(C+M+Y+K) is the minimum amount of C+M+Y+K obtained from the level set. The data in FIG. 6 for the original CMY value {0.7, 1, 1} results in the total amount of toner/ink being 270%. The minimum total amount of toners/inks in the CMYK level set is 84.2% corresponding the CMYK value of {0.066186, 0.031126, 0, 0.74461} resulting in a total toner savings of 185.8%. Because values in a CMYK level set correspond to the same CIELab value, the toner savings is achieved with little or no quality degradation.

Color Conversion with Ink Limitations

Another challenging task in color printing relates to the limitation of the amount of toner/ink coverage on a page to avoid excessive bleed-through. This task is particularly important for inkjet printing. When there exists a limit for the maximum amount of toner/ink coverage allowed per spot color printing, color conversion in highly saturated color regions or dark neutral color regions are generally affected. The color conversion affected by this limit includes the CIELab to CMY color conversion. The L₁ norm approximation may provide a suitable solution for solving the neutral color conversion problem and the L_∞ approximation may be suitable for color conversion in saturated color regions. For both approaches, RBFs with linear programming may be employed. The existence of a toner/ink limitation may be added as a constraint in the linear programming problem formulations. For example, the color conversion with the toner/ink limitation problem may now be formulated via both L₁ and L_∞ approximation algorithms. The L₁ algorithm may be a modified BR primal algorithm with an added constraint while the L_∞ algorithm may be developed based on the BP dual algorithm to extend the three-stage algorithm to a four-stage algorithm.

To analyze ink limitation problem using the L₁ approximation, g(z) is used to represent the mapping from CIELab to CMY in the form of an RBF as:

$\begin{array}{cc}{g}_k\left(z_j\right)=A_kz_j+{\alpha }_{k0}+\sum _{i=1}^n{\alpha }_{\mathrm{ki}}{\phi }_k\left(z_j-c_i\right),& \left(\mathrm{Eq}.163\right)\end{array}$

where k=1, 2, 3 corresponding to {g_C, g_M, g_Y}. Then, z_j={L, a, b}_j, where i=1, 2, . . . , n and j=1, 2, . . . , m. The dimension of A_k is 1×3 corresponding to the dimension of the domain. The dimension of a₀ is 1. Therefore, there are n+4 coefficients in such that Eq. 163 may be written as:

$\begin{array}{cc}{g}_k\left(z_j\right)=\underset{\underset{F_k\left(z_j\right)}{}}{\left(z_j,1,{\phi }_k\left(z_j-c_1\right),{\phi }_k\left(z_j-c_2\right),\dots ,{\phi }_k\left(z_j-c_i\right)\right)}\underset{\underset{w_k}{}}{\left(\begin{array}{c}{A}_k^T\\ a_{k0}\\ a_{k1}\\ {\alpha }_{k2}\\ \dots \\ {\alpha }_{\mathrm{kn}}\end{array}\right)n+4}& \left(\mathrm{Eq}.164\right)\\ =\sum _{i=1}^{n+4}w_{\mathrm{ki}}F_{\mathrm{ki}}\left(z_i\right).& \left(\mathrm{Eq}.165\right)\end{array}$

The L₁ error for each data point of component of C, M and Y using nonnegative variables are defined as

$\begin{array}{cc}{g}_k\left(z_i\right)-\sum _{i=1}^{n+4}w_{\mathrm{ki}}F_{\mathrm{ki}}\left(z_i\right)=u_{\mathrm{kj}}-v_{\mathrm{kj}}& \left(\mathrm{Eq}.166\right)\end{array}$

and the coefficients for each component using nonnegative variables are defined as w_ki=b_ki−c_ki, where u_kj, v_kj, b_ki, c_ki≧0. The primal linear programming problem in the L₁ sense may be illustrated by the constraint of the toner/ink limit applying to the color data with a toner/ink summation exceeding the total amount of toner or ink (e.g., C+M+Y+K>L, where L is the total amount of toner or ink). In this regard,

$\begin{array}{cc}\mathrm{Minimize}\xi =\sum _{j=1}^mu_{1j}+v_{1j}+u_{2j}+v_{2j}+u_{3j}+v_{3j}\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.167\right)\\ \sum _{i=1}^{n+4}\left(a_{1j}-b_{1j}\right)F_{1j}\left(z_j\right)+u_{1j}+v_{1j}=C_j& \left(\mathrm{Eq}.168\right)\\ \sum _{i=1}^{n+4}\left(a_{2j}-b_{2j}\right)F_{2j}\left(z_j\right)+u_{2j}+v_{2j}=M_j& \left(\mathrm{Eq}.169\right)\\ \sum _{i=1}^{n+4}\left(a_{3j}-b_{3j}\right)F_{3j}\left(z_j\right)+u_3+v_{3j}=Y_j& \left(\mathrm{Eq}.170\right)\\ \sum _{i=1}^{n+4}\left(a_{1j}-b_{1j}\right)F_{1j}\left(z_{\mathrm{jl}}\right)+\sum _{i=1}^{n+4}\left(a_{2j}-b_{2j}\right)F_{2j}\left(z_{\mathrm{jl}}\right)+\sum _{i=1}^{n+4}\left(a_{3j}-b_{3j}\right)F_{3j}\left(z_{\mathrm{jl}}\right)\le L,\mathrm{where}J=\left\{j_l|C_{\mathrm{jl}}+M_{\mathrm{jl}}+Y_{\mathrm{jl}}>L\right\}\subseteq \left\{i=1,2,\dots ,m\right\},u_{\mathrm{kj}},v_{\mathrm{kj}},b_{\mathrm{ki}},c_{\mathrm{ki}}\ge 0,k=1,2,3,i=1,2,\dots ,n,\mathrm{and}j=1,2,\dots ,m.& \left(\mathrm{Eq}.171\right)\end{array}$

With this in mind, a condensed initial simplex table for the CIELab to CMY conversion with the toner/ink limitation L is established in the table below where p denotes the number of color data points that exceed the toner/ink limit. The slack variables si's are then added in Eq. 171.

The slack variables s_i's are then added in Eq. 171.

The modified BR's algorithm may then be described in two stages. In the first stage, there are three sets of n iterations for the C, M, and Y components respectively. In each of the n iterations, the pivotal columns are restricted to b_i and c_i. The vector entering the basis is the one with the largest nonnegative marginal cost. The vector leaving the basis is chosen from v_j and u_j for the one causing the maximum reduction in the objective function. The interchange of vectors leaving and entering the basis are these performed within the same color component.

In the next stage, the nonbasic u_j or v_j are interchanged with the basic u_j, v_j, or s_jl. Neither b_i or c_i is allowed to leave the basis. The criteria for selecting vectors leaving and entering the basis are the same as that in the first stage. The interchange of vectors v_j and u_j leaving and entering the basis may be performed within the same color component. However, wj_l can be interchanged with the v_j and u_j vector for any color component. Once the vector s_jl leaves the basis, it is generally not allowed to enter back onto the basis. The iteration stops when all the marginal costs are non-positive. Thereafter, basic vectors b_i or c_i may be interchanged with the corresponding nonbasic vectors c_j or b_j.

As discussed above, the BP dual algorithm may be a relatively good algorithm for color conversion using L_∞ approximation. For the color conversion with the toner/ink limitation, a constraint may be added in the problem formulation of the BP dual algorithm using the L_∞ approximation.

In this regard, the L_∞ errors may be defined for component of C, M and Y:

$\begin{array}{cc}{\xi }_c=\underset{1\le j\le m}{\max }{\sum _{i=1}^{n+4}w_{\mathrm{ci}}F_{\mathrm{ci}}\left(z_j\right)-C_j}_{\infty }& \left(\mathrm{Eq}.172\right)\\ {\xi }_m=\underset{1\le j\le m}{\max }{\sum _{i=1}^{n+4}w_{mi}F_{mi}\left(z_j\right)-M_j}_{\infty }& \left(\mathrm{Eq}.173\right)\\ {\xi }_y=\underset{1\le j\le m}{\max }{\sum _{i=1}^{n+4}w_{\mathrm{yi}}F_{\mathrm{yi}}\left(z_j\right)-Y_j}_{\infty }& \left(\mathrm{Eq}.174\right)\end{array}$

Then, the problem may be formulated in the primal form by minimizing ξ=ξ_c+ξ_m+ξ_y, subject to

$\begin{array}{cc}{\xi }_c+\sum _{i=1}^{n+4}{\alpha }_{\mathrm{ci}}F_{\mathrm{ci}}\left(z_j\right)\ge C_j& \left(\mathrm{Eq}.175\right)\\ {\xi }_c-\sum _{i=1}^{n+4}{\alpha }_{\mathrm{ci}}F_{\mathrm{ci}}\left(z_j\right)\ge -C_j& \left(\mathrm{Eq}.176\right)\\ {\xi }_m+\sum _{i=1}^{n-4}{\alpha }_{mi}F_{mi}\left(z_j\right)\ge M_j& \left(\mathrm{Eq}.177\right)\\ {\xi }_m-\sum _{i=1}^{n+4}{\alpha }_{mi}F_{mi}\left(z_j\right)\ge -M_j& \left(\mathrm{Eq}.178\right)\\ {\xi }_y+\sum _{i=1}^{n-4}{\alpha }_{\mathrm{yi}}F_{\mathrm{yi}}\left(z_j\right)\ge Y_j& \left(\mathrm{Eq}.179\right)\\ {\xi }_y-\sum _{i=1}^{n+4}{\alpha }_{\mathrm{yi}}F_{\mathrm{yi}}\left(z_j\right)\ge -Y_j& \left(\mathrm{Eq}.180\right)\\ -\sum _{i=1}^{n+4}\left({\alpha }_{\mathrm{ci}}F_{\mathrm{ci}}\left(z_{\mathrm{jl}}\right)+{\alpha }_{mi}F_{mi}\left(z_{\mathrm{jl}}\right)+{\alpha }_{\mathrm{yi}}F_{\mathrm{yi}}\left(z_{\mathrm{jl}}\right)\right)\le -L,& \left(\mathrm{Eq}.181\right)\end{array}$

where ξc, ξm, ξy≧0, a₁, a₂, . . . a_n+4 unrestricted, z_j={L, a, b}_j, j=1, 2, . . . , m, J={jl|C_jl+M_jl+Y_jl>L}⊂{j=1,2, . . . , m} and F_j=[l, z_j, φ₁(z_j), φ₂(z_j), . . . , φ_n(z_j)]_n+4.

Transformed from the primal formulation, the dual formulation is performed by maximizing

$\begin{array}{cc}\sum _{i=1}^m\left[\left({\underset{\_}{\sigma }}_{\mathrm{ci}}-{\underset{\_}{\tau }}_{\mathrm{ci}}\right)C_i+\left({\underset{\_}{\sigma }}_{mi}-{\underset{\_}{\tau }}_{mi}\right)M_i+\left({\underset{\_}{\sigma }}_{\mathrm{yi}}-{\underset{\_}{\tau }}_{\mathrm{yi}}\right)Y_i\right]-L\sum _{i_k=1}^p{\underset{\_}{\rho }}_{i_k},\mathrm{subject}\mathrm{to}& \left(\mathrm{Eq}.182\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{cj}}-{\underset{\_}{\tau }}_{\mathrm{cj}}\right)F_{c_{\mathrm{ij}}}^T-\sum _{\mathrm{jl}=1}^p{\underset{\_}{\rho }}_jF_{c_{{\mathrm{ij}}_l}}^T=0& \left(\mathrm{Eq}.183\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{cj}}+{\underset{\_}{\tau }}_{\mathrm{cj}}\right)\le 1& \left(\mathrm{Eq}.184\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{mj}}-{\underset{\_}{\tau }}_{\mathrm{mj}}\right)F_{m_{\mathrm{ij}}}^T-\sum _{\mathrm{jl}=1}^p{\underset{\_}{\rho }}_jF_{m_{{\mathrm{ij}}_l}}^T=0& \left(\mathrm{Eq}.185\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{mj}}+{\underset{\_}{\tau }}_{\mathrm{mj}}\right)\le 1& \left(\mathrm{Eq}.186\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{yj}}-{\underset{\_}{\tau }}_{\mathrm{yj}}\right)F_{y_{\mathrm{ij}}}^T-\sum _{\mathrm{jl}=1}^p{\underset{\_}{\rho }}_jF_{y_{{\mathrm{ij}}_l}}^T=0,& \left(\mathrm{Eq}.187\right)\\ \sum _{j=1}^m[\left({\underset{\_}{\sigma }}_{\mathrm{yj}}+{\underset{\_}{\tau }}_{\mathrm{yj}}\right)\le 1,\mathrm{and}& \left(\mathrm{Eq}.188\right)\\ {\underset{\_}{\sigma }}_j,{\underset{\_}{\tau }}_j,{\underset{\_}{\rho }}_j\ge 0,j=1,2,\dots ,m,& \left(\mathrm{Eq}.189\right)\end{array}$

where z_j={L, a, b}_j, j=1, 2, . . . , m, J={jl|C_jl+M_jl+Y_jl>L}⊂{j=1,2, . . . ,m}, F_j32 [l, z_j, φ_l(z_j), φ₂(z_j), . . , φ_n (z_j)]_n+4. F_jl=[l, z_jl, φ_jl (z_jl), φ_j2 (z_jl), . . . , φ_n(z_jl)]_n+4, φ_ij=φ_i(z_j), and p denotes the number of color data points that exceed the toner/ink limit.

A modified BP L_∞ approximation algorithm may be used for solving the color conversion from CIELab to CMY with a toner/ink limitation L based on the condensed initial simplex table below.

In the first stage of the BP algorithm, the first n simplex iterations are performed. For example, the a_c vectors may be moved out of the basis. The vector entering the basis is selected from s_c to correspond to that with the largest absolute reduced cost. The a_m vectors may be moved out of the basis. The vector entering the basis is selected from s_m to correspond to that with the largest absolute reduced cost. The a_y vectors may be moved out of the basis. The vector entering the basis is selected from s_y to correspond to that with the largest absolute reduced cost.

In the second stage of the BP algorithm, ξ_c is forced out of the basis. The vector entering the basis is selected from the remaining nonbasic vectors s_c and a_c is not allowed back onto the basis. ξ_m is forced out of the basis. The vector entering the basis is selected from the remaining nonbasic vector s_m and a_m is not allowed to move back onto the basis. ξ_y is forced out of the basis. The vector entering the basis is selected from the remaining nonbasic vector s_y and a_y is not allowed to move back onto the basis.

In the third stage of the BP algorithm, the pivotal column is chosen corresponding to the most negative marginal cost. The pivotal row may then be chosen by the ratio selection rule. The interchange of nonbasic vectors s_i and basic vectors s_j may be performed within the same color component. The iteration continues until all marginal costs corresponding to the nonbasic vectors s_c, s_m, and s_y, are nonnegative. The vector s_cmy is not allowed to enter the basis.

A fourth stage is added to perform simplex iteration interchanging nonbasic vectors including all the remaining vectors s_c, s_m, s_y, and s_cmy and the basic vectors in the basis. Nonbasic vectors s_c, s_m, and s_y interchange with basic vectors of s_c, s_m, and s_y respectively. The nonbasic vectors s_c, s_m, and s_y interchange with any basic vectors of s_c, s_m, and s_y ending when the marginal costs are nonnegative. The function φ thus satisfies the Haar condition as described in above.

Gamut Mapping

The inventive embodiments of out-of-gamut conversion that include CIELab to CMYK gamut conversion may be described as follows. First, the thresholds for color difference in hue angle Th and color difference in lightness TL based on the perceptual color difference tolerance are defined. Next, the output device gamut boundary gmt0 specified in the CIELab color space is defined. This boundary may be expanded into a few layers with a small increment each time (e.g., {gmt_i, i=1, 2, . . . , N}). In each layer, the CIELab values are {(L, a, b)_ij, j=1, 2, . . . ,m_i}. The corresponding color in CIELch is {(L, c, h)_ij, j=1, 2, . . . ,m_i} where mi is the number of the CIELab colors in the layer i. Then the CMY value is obtained using

g₀((L, a, b)_ij)−(C,M,Y)_ij,   (Eq. 190)

where g₀ is the mapping function from CIELab to CMY based on the inside gamut colors. The CIELab data in the first expanded layer is (L, a, b)_i={(L, a, b)_ij, j=1,2, . . . ,m₁}. The number of color data points in gmt1 is (C,M,Y)₁={(C,M,Y)_ij, j=1,2, . . . ,m₁}, m₁. If any C, M, and Y component value in (C, M, Y) 1 is smaller than 0 or greater than 1, then the value is changed to 0 or 1 respectively. The (C, M, Y )1 may then be changed to (C, M, Y)′1. (L, a, b)′1 is obtained using

f₀(C′_1j,M′_1j,Y′_1j,0)→(L′_1j,a′_1j,b′_1j)   (Eq. 191)

where f₀ is the mapping function from CMYK to CIELab based on the inside gamut colors. Then (L, a, b)′1 to (L, C, h)′1 is converted and the lightness difference ({dL_1j,j=1, 2, . . . ,m₁} and the hue difference ({dh_1j, j=1, 2, . . . ,m₁} are calculated for each data point between (L, C, h)₁ and (L, C, h)′₁. If dh_1j>T_h or dLlj>T_L, (L, C, h)′₁ is converted to (L, C, h)″₁ by making the following changes in L′_1j and h′_1j:

$\begin{array}{cc}L=\{\begin{array}{c}{L}_{1,j}+T_L,L_{1j}^{\prime }>L_{1j}\\ L_{1j}-T_L,L_{1j}^{\prime }<L_{1j}\end{array}& \left(\mathrm{Eq}.192\right)\\ h=\{\begin{array}{c}{h}_{1j}+T_h,h_{1j}^{\prime }>h_{1j}\\ h_{1j}-T_h,h_{1j}^{\prime }<h_{1j}.\end{array}& \left(\mathrm{Eq}.193\right)\end{array}$

The steps may be repeated until the (C, M, Y)₁ values are found to meet the criteria of T_h and T_L. Then all out-of-gamut CIELab values (L, a, b)₁ may be mapped in gmt1 with (C, M, Y)₁. Models of g₁ and f₁ are then generated using the data points in gmt1:

g₁((L,a,b)₁)→(C,M,Y) and   (Eq. 194)

f₁((C,M,Y)₁)→(L, a, b).   (Eq. 195)

Each of the above steps may then be repeated and the CIELab data interpolated in {gmt_i, i=2,3, . . . ,N)} to find {(C,M,Y)_ij, j=1,2, . . . ,m_i}. The iteration stops when the last color layer reaches optimized color ranges (e.g., as described in U.S. Pat. No. 6,719,392, the entire contents of which are incorporated by reference)

When the RBF models f_i and g_i are generated based on the data points in the gamut layer gmt_i, the data used in generating these models may contain two parts. The first part includes the data that fits the model g_(i−1) and the data of f_(i−1) that is generated based on the gamut layer i−1. The second part includes the data not fitting g_(i−1) and f_(i−1). This second part has C, M, or Y values either below 0 or above 1 using f_(i−1) and is clipped to either 0 and 1 and refined with the thresholds T_h and T_L.

In doing so, the models generated for each gamut layer are continuous and smooth. To summarize, the gamut mapping direction is obtained by creating the numerical model guided by a perceptual color difference model. The advantages of this technique are that more levels are obtained from this gamut mapping algorithm. Moreover, out-of-gamut colors in a close neighborhood vary continuously. The direction of the gamut mapping is not fixed, rather it varies smoothly to provide aesthetically pleasing color mappings.

FIG. 7 is a flowchart illustrating a method 700 of gamut mapping in one exemplary embodiment of the invention. The method 700 initiates when the processor 110 determines the color gamut of an imaging device (e.g., the digital camera 116, the computer monitor 112, the printer 106, or the like) in step 701. The processor 110 may then receive color data having color values that lie outside of the color gamut of the imaging device in step 702. For example, the processor 110 may receive color data from the digital camera 116 that is to be printed on the printer 106. The digital camera 116 may have an input color gamut that is larger than the output color gamut of the printer 106. That is, the digital camera may capture images with color values that lie outside the output color gamut of the printer 106. Accordingly, the colors of the digital camera 116 are mapped to the output color gamut of the printer 106 using the gamut mapping described above. In this regard, the processor 116 may determine a color region that is external to the output color gamut of the imaging device in step 703 and map color values from that color region to the color gamut of the imaging device in step 704. To map the colors, the processor 110 may employ the numerical mapping using RBFs as described herein.

Once the colors of the region are mapped to the color gamut of the imaging device, the processor 116 may remap the color values to the region from the output color gamut in step 705. Thereafter, the processor 116 may compare the remapped color values to the original color values to determine whether the mapping is aesthetically pleasing. In this regard, the processor 116 may determine whether the remapped color values are proximate to the original color values in step 707. If the remapped color values are not proximate to the original color values, color values from a perceptual model may be selected as suitable substitutes for the color values of the input color gamut in step 708. Generally, comparisons of color values are performed individually such that suitable color values from the perceptual model are also individually selected. That is, entire regions of color values are generally not substituted with color values from a perceptual module for the failure of a single color value to map to the output color gamut in an aesthetically pleasing manner.

If the remapped color values are proximate to the original color values, the mapping of color values may continue in step 709 where the processor 116 determines whether all of the color values of the color region have been mapped. If all of the color values of the color region are mapped to the output color gamut (e.g., via gamut mapping or perceptual color mapping), the gamut mapping process concludes in step 710. If the color values of the color region have not been completely mapped to the output color gamut, the area of the color region may be increased in step 711. That is, another layer surrounding the previous color layer is determined in a manner similar to step 703 such that color values within the expanded color region may be mapped to the output color gamut in step 712. From there, the method 700 returns to step 705 to ensure the gamut mapping is fairly accurate. Thus, the method 700 continues until each of the color values from the input color gamut from mapped to the output color gamut.

Although specific embodiments were described herein, the scope of the invention is not limited to those specific embodiments. The scope of the invention is defined by the following claims and any equivalents thereof. For example, the context of the detailed description often relates to gamut mapping from an imaging device, such as a camera, for the purposes of printing in a different, typically smaller color gamut. However, the invention is not intended to be so limited. Rather, the color gamut mapping described herein may be used to map from virtually any color gamut to another. For example, color gamut of a digital camera may be mapped to the color gamut of a computer monitor. In another example, a scanned image having one color gamut may be converted to the color gamut of another device such as a digital camera, a computer monitor, or a printer, for display therewith.

In the context of printing, the invention is not intended to be limited to any particular type of printer. For example, the gamut mapping described herein may be employed in either ink or toner based printers. Moreover, the gamut mapping may be employed in printers ranging in size from small household printers to large commercial printers. In one embodiment, the gamut mapping may be implemented as software instructions operable with a host system to map colors between various gamuts as needed. For example, software instructions may direct a processing system, such as a general purpose computer, to map the gamut of one device to the gamut of another device prior to presentation to the other device.

Claims

1. A method of color conversion, the method comprising: determining a color gamut of an imaging device;receiving image data having color values external to the color gamut of the imaging device;determining a color region of the image data external to the color gamut;mapping the color values within the color region to the color gamut;iteratively increasing an area of the color region; andmapping the color values within the increased color region to the color gamut.

2. The method of claim 1, further comprising generating a numerical model for mapping the color values within the color region to the color gamut.

3. The method of claim 2, wherein generating the numerical model includes generating a radial basis function.

4. The method of claim 1, further comprising: generating a numerical model for mapping mapped color values to the color region; andmapping a mapped color value to the color region based on the numerical model to determine whether the mapped color value is proximate to an original color value.

5. The method of claim 4, wherein mapping a mapped color value to the color region based on the numerical model comprises mapping a CMYK mapped color value to a CIELab color region.

6. The method of claim 4, further comprising comparing the remapped color value to a perceptual color model to determine whether the mapped color value is proximate to a human visual system value.

7. The method of claim 1, wherein mapping the color region to the color gamut comprises mapping a CIELab color region to a CMYK color gamut.

8. The method of claim 1, wherein iteratively increasing an area of the color region comprises expanding the area of the color region in color region increments based on a numerical model.

9. A color conversion system, comprising: an imaging device operable to generate image data having a plurality of color values; anda processor operable to convert the color values of the image data to a color gamut of another imaging device, wherein the processor determines a color region external to the color gamut, maps the color values within the color region to the color gamut, iteratively increases an area of the color region, and maps the color values within the increased color region to the color gamut.

10. The color conversion system of claim 9, wherein the processor is further operable to generate a numerical model for mapping the color values within the color region to the color gamut.

11. The color conversion system of claim 10, wherein the numerical model includes generating a radial basis function.

12. The color conversion system of claim 9, wherein the processor is further operable to generate a numerical model for mapping mapped color values to the color region and to map a mapped color value to the color region based on the numerical model to determine whether the mapped color value is proximate to an original color value.

13. The color conversion system of claim 12, wherein processor is further operable to map a CMYK mapped color value to a CIELab color region based on the numerical model.

14. The color conversion system of claim 12, wherein processor is further operable to compare the remapped color value to a perceptual color model to determine whether the mapped color value is proximate to a human visual system value.

15. The color conversion system of claim 9, wherein processor is further operable to map a CIELab color region to a CMYK color gamut.

16. The color conversion system of claim 9, wherein the area of the color region comprises color region increments based on a numerical model.

17. A printing system, comprising: a printer;an imaging device operable to generate image data having a plurality of color values; anda printer controller operable to process the color values of the image data and map the color values to a color gamut of the printer by determining a color region external to the color gamut, mapping the color values within the color region to the color gamut, increasing an area of the color region, and mapping the color values within the increased color region to the color gamut.

18. A method of color conversion, the method comprising: determining a color gamut of an imaging device;receiving image data having color values external to the color gamut of the imaging device;determining a color region of the image data external to the color gamut;mapping the color values within the color region to the color gamut via a numerical model of the color gamut;mapping the mapped color values to the color region via a numerical model of the color region; andcomparing the mapped color values of the color region to original color values of the color region to determine whether the mapped color values correspond to the original color values.

19. A software program used in color conversion, the software program comprising instructions that are configured to direct a processor to: determine a color gamut of an imaging device;receive image data having color values external to the color gamut of the imaging device;determine a color region of the image data external to the color gamut;map the color values within the color region to the color gamut;iteratively increase an area of the color region; andmap the color values within the increased color region to the color gamut.