This application claims the benefit under 35 U.S.C. §119(a) to a Singapore patent application filed in the Intellectual Property Office of Singapore on Dec. 31, 2008 and assigned Serial No. 200809685-1, the entire disclosure of which is hereby incorporated by reference.
The present invention generally relates to the field of digital image processing, and more particularly to a system and process for rescaling digital images using adaptive interpolation kernel with sharpness and overshoot control.
Digital images have become more popular in the field of image display because they offer clarity and less distortion during processing. Furthermore, a wider range of image processing algorithms can be applied to digital images. Interpolation is a common stage in image processing to improve the appearance of the processed image on the output imaging medium. Interpolation is often performed during rescaling or resizing of digital images.
Rescaling or resizing of digital images includes magnification or reduction of image. For example, large screen displays have a native resolution that reaches or exceeds the well-known high-definition TV (HDTV) standard. In order to display a low-resolution digital image on a large screen display, it is desirable to rescale the image to a full screen resolution.
Traditionally, linear interpolation techniques such as bilinear or bicubic interpolation are used to rescale digital images. However, these linear interpolation techniques have certain drawbacks. For example, the interpolated image may suffer from blurred edges, resulting in a loss in image quality. Furthermore, artifacts such as overshoot and undershoot may be introduced around the edge areas. Therefore, many techniques have been proposed to improve the subjective quality of the interpolated images by using adaptive (interpolation) algorithms. These adaptive algorithms minimize loss in image quality by adapting the interpolation coefficients in the algorithms to better match the local structures around the edges of the image.
One approach/method for adaptive interpolation of image data is by selecting an appropriate interpolation kernel function from a plurality of different interpolation kernel functions based on one or more factors. The factors can be the input image format, the output imaging media, or the appearance characteristic of the interpolated image data.
Another approach/method (to provide a representation of image data) is by using a scaled kernel. The scaled kernel is constructed by transforming a kernel from a first range to a second range, where the second range is smaller than the first range. This method can improve the sharpness of the image data without any overshoot problem. However, transforming a kernel from one range to another range creates a dead-zone in the transformed range, which may result in an un-natural edge effect in the interpolated image.
Yet another approach/method (for digital image interpolation and sharpness enhancement) is by combining two interpolated images with sharpness level control. The two interpolated images are generated by two interpolation filters with different cut-off frequencies. This method will lead to overshoot and undershoot artifacts near the edge areas when using large sharpness control parameters.
Generally, the three methods as discussed above address the issues of loss in image quality by applying effective global sharpness control. In global sharpness control, the control parameter is fixed for all the pixels. However, the effect of local sharpness control for different characteristics of the image content has not been adequately addressed. For example, while the three methods enhance the sharpness of the edge components, they also tend to amplify the noise in the signal. Furthermore, the three methods are unable to control the sharpness level locally through local sharpness control, thereby causing the weak edges to be enhanced without over-enhancing the strong edges. In local sharpness control, the control parameter is different for each individual pixel. As a result, the methods produce artificial-looking images.
One approach for local sharpness control is to apply an adaptive enhancement process on different edge magnitudes to achieve good edge enhancement. However, the soft edges can add depth perceptibility to the digital image, and therefore this approach is not suitable for steepening. Furthermore, the hard edges are already steep enough to demand further enhancement.
As described above, conventional linear interpolation techniques and the existing adaptive techniques suffer from blurred edges or artifacts introduced around edge areas, such as overshoot and undershoot. The conventional techniques generally address the effectiveness of global sharpness control which is based only on different formats of input image data or different types of the output display media. They are not concerned about the local image content. Hence, they are noise sensitive.
In the case of downscaling, a number of input pixels in the original image are transformed to a single output pixel in the down-sampled image. The frequency sampling interval must be chosen very short based on Nyquist sampling theory in order to transform the entire data from input to output image, which is not feasible in most cases, and aliasing artifacts can be introduced.
As shown in
Preferred embodiments according to the present invention will now be described with reference to the Figures, in which like reference numerals denote like elements.
The present invention may be understood more readily by reference to the following detailed description of certain embodiments of the invention.
The present invention provides an effective process of video image rescaling. This process uses a controllable interpolation kernel to generate the interpolated outputs ranging from relatively sharp at one end to relatively soft at the other end. The levels of sharpness and overshoot can be either selected by users using a set of control parameters or adaptively steered by the image contents. In addition, a downscaling process with anti-aliasing low pass filtering is used to generate the downsampled images without aliasing artifacts.
The present invention provides a system for video image rescaling using a controllable interpolation kernel. The system comprises a global and local sharpness control functionality to generate the interpolated images with the visual qualities ranging from relative sharpness to softness. Furthermore, the system comprises an overshoot control functionality to adjust the levels of overshoot/undershoot effects along the edges in the interpolated images. In addition, a set of anti-aliasing low pass filters are designed and conducted according to a certain range of downscaling factors.
Now referring to
The adaptive spline interpolation kernel 101 uses a controllable third order polynomial function based on four input support pixels received from the input image data to generate an interpolated pixel. One embodiment of the implementation of the adaptive spline interpolation kernel 101 is shown in
where yA, yB, yC and yD are four support pixels from input image data; s is the distance from interpolation position to the position of the support pixel yb as shown in
f
A(s,Kshp)=Kshp*(−s3+2*s2−s)/2, [Eqn. 2]
f
B(s,Kshp)=(3*Kshp*s3−5*Kshp*s2+2*(Kshp−1)*s+2)/2, [Eqn. 3]
f
C(s,Kshp)=(−3*Kshp*s3+4*Kshp*s2+(2−(Kshp)*s)/2, and [Eqn. 4]
f
D(s,Kshp)=Kshp*(s3−s2)/2. [Eqn. 5]
Kshp is the global sharpness control parameter generated from the sharpness control module 102, wherein Kshp is used to adjust the sharpness of the rescaled image. Kshp has a predetermined range of negative to positive values to control the visual quality of the rescaled images from soft to sharp, wherein a negative Kshp produces a softer rescaled image and a positive Kshp produces a sharper rescaled image.
From
Referring back to
The outputs y1(x) and y2(x) are then subtracted to produce an output (y1(x)−y2(x)). The multiplier 204 multiplies the output (y1(x)−y2(x)) by the overshoot control parameter Kovs to produce the output Kovs*(y1(x)−y2(x)). Thereafter, Kovs*(y1(x)−y2(x)) is added to y2(x) to generate the final output y(x) with sharpness and overshoot control according to the following Equation 7:
y(x)=Kovs*(y1(x)−y2(x))+y2(x). [Eqn. 7]
An alternative embodiment of the adaptive spline interpolation kernel with sharpness control 201 is shown in
f
A(s)=(−s3+2*s2−s)/2, [Eqn. 8]
f
B(s)=(3*s3−5*s2+2)/2, [Eqn. 9]
f
C=(−3*s3+4*s2+s)/2, and [Eqn. 10]
f
D(s)=(s3−s2)/2. [Eqn. 11]
The size of this look up table is dependent on the resolution of interpolation, which is initialized at the beginning of the rescaling process. The spline interpolation module 602 is used to calculate the interpolated pixel value yc(X) using the KernelLUT 601. The interpolation can be obtained according to the following Equation 12:
y
c(x)=y(A)*fA(s)+y(B)*fB(s)+y(C)*fC(s)+y(D)*fD(s). [Eqn. 12]
The linear interpolation 603 is used to calculate the interpolated pixel value yt(x) which can be obtained according to the following Equation 13:
y
t(x)=(1−s)*y(B)+s*y(C). [Eqn. 13]
The outputs yc and yt(x) are then subtracted to produce an output (yc(x)−yt(x)). The multiplier 604 multiplies the output (yc(x)−yt(x)) by the sharpness control parameter Kshp to produce the output Kshp*(yc(x)−yt(x)). Thereafter, Kshp*(yc(x)−yt(x)) is added to yt(x) to generate the final output y1(x) with sharpness control according to the following Equation 14:
y
1(x)=Kshp*(yc(x)−yt(x))+yt(x). [Eqn. 14]
Referring back to
The local gradient estimation module 104 is coupled to the input module 108 to estimate the local gradient magnitude Gn based on a 3×3 neighbourhood window (shown in
G
n=(abs(X0+2*X1+X2−(X6+2*X7+X8))+abs(X0+2*X3+X6−(X2+2*X5+X8)))/8. [Eqn. 15]
The noise level estimator module 107 is coupled to the input module 108 to estimate the noise level Sigma from the input image or video. An example of a suitable noise level estimator or detector is disclosed in U.S. Pat. No. 5,657,401, the content of which is incorporated herein by reference. Alternatively, the noise level can be predetermined before transmitting it to the system so as to simplify the implementation of the system.
The local gradient estimation module 104 and the noise level estimator module 107 are both coupled to the local sharpness control parameter function module 105, wherein the local gradient estimation module 104 transmits the local gradient magnitude Gn to the local sharpness control parameter function module 105, and wherein the noise level estimator module 107 transmits the noise level Sigma to the local sharpness control parameter function module 105. The local sharpness control parameter function module 105 is configured to control the sharpness control module 102 using a transfer function, wherein the transfer function indicates the relationship between the control parameter Kshp and the local image content.
Kshp(Gn)=AKshp(Gn)+GKshp. [Eqn. 16]
GKshp is a fixed value that is defined by the global sharpness control parameter Kshp. AKshp(Gn) is a variable value that is defined by the control parameters max K, min K, K1, K2, K3, and Sigma. The AKshp(Gn) can be expressed by a transfer function according to the following Equation 17:
When the local gradient magnitude value (Gn) is lower than the noise level Sigma, the pixels are considered as being corrupted by noise. In this case, the sharpness control value will be lower than GKshp and decreases linearly to a minimum value (min K+GKshp), as illustrated by segment 801 in
Referring back to
The anti-aliasing filter 106 comprises a plurality of 1 D low pass filters that are activated according to a predetermined range of downscaling factors.
where m is the length of the filter.
The low pass filter coefficients stored in the low pass filter bank 902 are designed using finite impulse response (FIR) filter design methods based on adjustable Hanning window. The cutoff frequencies used in the filter design depend on the downscaling factors. Therefore, different downscaling factors correspond to different low pass filters. In one embodiment, eight filters are operated according to eight pre-divided ranges of downscaling factors.
While this detailed description has set forth some embodiments of the present invention, the appended claims are sufficiently supported to cover and will cover other embodiments of the present invention which differ from the described embodiments according to various modifications and improvements.
Number | Date | Country | Kind |
---|---|---|---|
200809685-1 | Dec 2008 | SG | national |