The present invention relates generally to the field of digital imaging and, more particularly, to an automatic focusing apparatus and method for digital images using automatic filter switching.
Facilitating the automatic capturing of high quality images with little or no user intervention is the primary objective in any consumer level digital or cell-phone camera system design. Consumers often rank image sharpness as a major factor of final image quality [1]. The sharpness of the final captured image is determined in large part by the ability of the camera's auto-focus (AF) system to adjust the distance between the lens and image sensor in such a way as to render an object of interest at a certain distance from the camera onto the image sensor with an imperceptible amount of blur.
Due to its ease of hardware integration, the passive AF approach has become the defacto standard in realizing the AF feature for compact consumer digital or cell-phone camera systems. In passive AF, a measure of image sharpness is extracted from a portion of the captured image. This measure is then used to adjust the imaging distance via a search algorithm running on the camera's processor so that the extracted measure obtains a peak value [2].
There are three components to any passive AF method namely: focus region, sharpness measure, and sharpness search algorithm [3]. As far as sharpness search algorithms are concerned, several attempts have been made to determine the in-focus position quickly without overshooting or oscillating around the peak as consumers desire a smooth AF experience with minimum overshoot and no rapid oscillation between extreme levels of focus and defocus [4], [5]. The efficiency of the search depends on the number of times the distance between the lens and image sensor is adjusted to bring the image into focus.
Many different types of search approaches can be found in the literature including Global Search, Iterative Coarse-to-Fine Search, Divide-and-Conquer Search, Prediction Search, and Variable Step-Size Sequential Search. The Global Search (GS) approach sequentially searches every position in the search space, and uses the position of maximum sharpness as the in-focus position. A GS-found in-focus position provides the true in-focus position and thus can be used to compare the accuracy of other search methods. Iterative coarse-to-fine approaches include Choi's fast hopping search [6], which combines a coarse scan with GS near the estimated peak location, and Li's fixed step-size coarse search [7], which uses a fine scan near the estimated coarse peak. The divide-and-conquer approach is exemplified with the Fibbonaci [8] search. Although this search is optimal in minimizing the number of iterations for a given search space, it is not a viable method for a consumer camera AF system due to the oscillation around the peak and its inefficiency in total distance moved [4], [5]. In [9], Chen et al. presented a prediction search to forecast the turning point of the sharpness measure which helped to reduce the number of iterations. Several variable step-size search methods have been proposed for adjusting the speed of search, the key differentiator in such methods lies in how to determine the step-size. Fuzzy rule-based [10], [11] and crisp rule-based [3], [12] methods have been applied to adjust the step-size in a heuristic manner, while in [13], Batten adapted the step-size in order to keep the gradient of the sharpness measure constant. In [14], Yao used a Maximum-Likelihood statistical approach to determine thresholds in a crisp rule-based type search for the adjustment of the step-size.
In [15], Yao et al. presented a study of sharpness measures and search algorithms, where it was found that the variable step-size Rule-based Search (RS) [12] outperformed other search algorithms in terms of accuracy and convergence speed. In other words, it has been independently verified that such a sequential search is able to reduce the number of iterations and to eliminate the oscillation found in the divide-and-conquer approaches. In [3], the effectiveness of the RS approach was also confirmed and improvements were made to it, named Modified-RS (MRS), that achieved better performance than the original RS.
The existing sharpness search algorithms presented in the literature, including RS and MRS, are performed based on user-defined thresholds, rules, and step-size magnitudes. However, since the sharpness measure depends on many factors including the optics, the sensor, the object of interest and its surrounding environment, it becomes difficult to tune such thresholds to control the adjustment of the step-size to achieve an efficient search across different camera platforms.
Furthermore, in the digital band-pass filter sharpness measure method that is popular with camera manufacturers, the choice of filter pass-band is somewhat arbitrary. Only general recommendations have been mentioned in the literature such as using a low-frequency band at the initial searching stage and then using a higher-frequency band to accurately locate the peak [16]. It is not mentioned how the pass-band should be changed when changing the camera optics and sensor.
Thus, from a camera manufacturer's point of view, it is desirable to parameterize the passive AF system so that it becomes portable across many different camera platforms. The existing AF algorithms, however, do not adequately address how to properly set the AF parameters, such as the filter pass-band and the step-size magnitude. Accordingly, there is a need for AF method and system that automatically derives the passive AF parameters, reduces the number of iterations and lowers the auto focusing time while not compromising sharpness quality.
The present invention provides an automatic focusing apparatus and method for digital images that automatically derives the passive AF parameters based on camera specifications, and uses the derived AF parameters in a Filter-Switching Search (FS) that reduces the number of iterations and lowers the auto focusing time while not compromising sharpness quality. The passive AF parameters are automatically derived based on the camera specifications including the lens F-number, the sensor's pixel pitch, and the focus actuator's step resolution. The derived parameters also include an optimum bank of digital band-pass filters matched to the camera specifications along with their corresponding step-size magnitudes.
The Filter-Switching Search (FS) takes advantage of the derived parameters to realize a self-tuning passive AF system where filters are switched and step-sizes are reduced as the search gets closer to the in-focus position by using local estimates of the first and second derivatives of the focus value surface. Such a formulation provides a means to automatically achieve a balance between AF speed and AF accuracy without extensive tuning or adjustment of difficult-to-set thresholds or parameters. In order to gauge the performance of passive AF systems, an experimental methodology is presented to standardize the performance testing of passive AF algorithms in terms of different light and contrast levels. By applying the developed framework on three different camera systems, it is demonstrated that the Filter-Switching Search (FS) outperforms the existing methods while requiring no or little parameter tuning/adjustment. The present invention can be used for AF systems in any digital imaging device, such as digital cameras (still or video), cell-phone cameras, computer cameras, personal data assistants (PDA), mobile communication devices, hand-held devices, vehicles, etc.
The present invention provides an automatic focusing apparatus that includes a focusing lens, a focus actuator connected to the focusing lens to move the focusing lens, an image sensor aligned with the focusing lens to detect an image through the focusing lens, and a memory storing a set of device parameters. The image sensor provides a noise-free, blurred, Bayer-pattern sampled image within a focus window from the detected image to the j digital band-pass filters connected to the image sensor to extract a set of sharpness samples from the noise-free, blurred, Bayer-pattern sampled image wherein j>=2. Furthermore, the apparatus includes a processor connected to the focus actuator, the digital band-pass filters and the memory. The processor (a) generates M sets of spatial frequency parameters and M focus actuator step sizes using the set of device parameters wherein M is a number of segments of a spatial frequency axis, and with i=1 (b) loads the i to i+j sets of spatial frequency parameters into the digital band-pass filters (c) obtains a matrix of focus value samples from the digital band-pass filters, (d) controls the focus actuator to sample a local surface behavior using the i focus actuator step size, (e) repeats steps (b) to (d) with i=i+1 whenever a switch condition is satisfied, and (f) estimates a final in-focus position for the image and controls the focusing lens to move to the final in-focus position whenever a stop condition is satisfied.
In addition, the present invention provides a method for automatically focusing a focusing lens using j digital band-pass filters wherein j>=2 by (a) generating M sets of spatial frequency parameters and M focus actuator step sizes using a set of device parameters wherein M is a number of segments of a spatial frequency axis, and with i=1, and (b) loading the i to i+j sets of spatial frequency parameters into the digital band-pass filters. Next, (c) a matrix of focus value samples is obtained from the digital band-pass filters, and (d) the focusing lens is controlled to sample a local surface behavior using the i focus actuator step size. Steps (b) to (d) are then repeated with i=i+1 whenever a switch condition is satisfied. Thereafter, (f) a final in-focus position is estimated for the image and the focusing lens controlled to move to the final in-focus position whenever a stop condition is satisfied. This method can be implemented by a computer readable medium encoded with a computer program wherein the steps are performed by one or more code segments.
Moreover, the present invention provides an automatic focusing apparatus and method in which M sets of spatial frequency parameters and M focus actuator step sizes are generated using the set of device parameters. The first and second sets of spatial frequency parameters are loaded into the first and second digital band-pass filters, respectively. The iteration k is initialized. If the stop condition is satisfied, the final in-focus position is estimated and the focusing lens is controlled to move to the final in-focus position. If, however, the stop condition is not satisfied, the sample state matrix is updated with the kth focus value samples. The samples in the state matrix are normalized for comparison of the focus values. If a switch condition is not satisfied, the process proceeds to the next iteration (k=k+1) and repeats the updating, normalization and switch condition check as previously described. If the switch condition is satisfied, local estimates of the second partial derivatives of the focus value surface are computed and the switch condition is computed. The switch matrix to be used in the next iteration is computed and the analysis filters to be used at the next iteration are computed. The sample state matrix is updated if the switch condition is satisfied. The step size is updated and the focus actuator position is updated. Thereafter, the process loops back to determine whether the stop condition is satisfied and repeats as previously described until the final in-focus position is estimated. When the final in-focus position is estimated, the focusing lens is controlled to move to the final in-focus position. This method can be implemented by a computer readable medium encoded with a computer program wherein the steps are performed by one or more code segments.
The present invention is described in detail below with reference to the accompanying drawings.
The above and further advantages of the invention may be better understood by referring to the following description in conjunction with the accompanying drawings, in which:
While the making and using of various embodiments of the present invention are discussed in detail below, it should be appreciated that the present invention provides many applicable inventive concepts that can be embodied in a wide variety of specific contexts. The specific embodiments discussed herein are merely illustrative of specific ways to make and use the invention and do not delimit the scope of the invention. The discussion herein relates primarily to digital cameras, but it will be understood that the concepts of the present invention are applicable to any digital image processing having automatic focusing.
The present invention provides an automatic focusing apparatus and method for digital images that automatically derives the passive AF parameters based on camera specifications, and uses the derived AF parameters in a Filter-Switching Search (FS) that reduces the number of iterations and lowers the auto focusing time while not compromising sharpness quality. The passive AF parameters are automatically derived based on the camera specifications including the lens F-number, the sensor's pixel pitch, and the focus actuator's step resolution. The derived parameters also include an optimum bank of digital band-pass filters matched to the camera specifications along with their corresponding step-size magnitudes.
The Filter-Switching Search (FS) takes advantage of the derived parameters to realize a self-tuning passive AF system where filters are switched and step-sizes are reduced as the search gets closer to the in-focus position by using local estimates of the first and second derivatives of the focus value surface. Such a formulation provides a means to automatically achieve a balance between AF speed and AF accuracy without extensive tuning or adjustment of difficult-to-set thresholds or parameters. In order to gauge the performance of passive AF systems, an experimental methodology is presented to standardize the performance testing of passive AF algorithms in terms of different light and contrast levels. By applying the developed framework on three different camera systems, it is demonstrated that the Filter-Switching Search (FS) outperforms the existing methods while requiring no or little parameter tuning/adjustment. The present invention can be used for AF systems in any digital imaging device, such as digital cameras (still or video), cell-phone cameras, computer cameras, personal data assistants (PDA), mobile communication devices, hand-held devices, vehicles, etc.
Now referring to
As will be described in more detail below, the processor 112 (a) generates M sets of spatial frequency parameters and M focus actuator step sizes using the set of device parameters wherein M is a number of segments of a spatial frequency axis, and with i=1 (b) loads the i to i+j sets of spatial frequency parameters into the digital band-pass filters 110 (c) obtains a matrix of focus value samples from the digital band-pass filters 110, (d) controls the focus actuator 104 to sample a local surface behavior using the i focus actuator step size, (e) repeats steps (b) to (d) with i=i+1 whenever a switch condition is satisfied, and (f) estimates a final in-focus position for the image and controls the focusing lens to move to the final in-focus position whenever a stop condition is satisfied. Moreover, the processor 112 generates the M sets of spatial frequency parameters and the M focus actuator step sizes by (a) producing a set of spatial frequencies along the spatial frequency axis using the set of device parameters, (b) determines an effective width in focus actuator steps for each spatial frequency using the set of device parameters, (c) segments the spatial frequency axis into a set of candidate band edges using the effective width for each spatial frequency, (d) constructs the M sets of spatial frequency parameters from the set of candidate band edges, and (e) computes the M focus actuator step sizes using the spatial frequency parameters and the set of device parameters.
Referring now to
This blur circle results in the well-known circular Point Spread Function (PSF),
The Fourier transform of the PSF is the Optical Transfer Function (OTF), and is given by [17]
HPSF(fsp;k)=2jinc(2πr(k)∥fsp∥2), (3)
where fsp=[f1f2]T is a vector of spatial frequencies in cycles per mm and
where J1(x) is the Bessel function of the first kind of order one. By substituting equation (1) in equation (3), one can derive the so-called through-focus OTF for the lens as follows [17]:
By fixing sO, fF, and F#, the through focus Modulation Transfer Function (MTF), |HPSF|, can be plotted to observe its behavior as a function of defocus amount ΔvI(k) and spatial frequency ∥fsp∥2 as shown in
For a given spatial frequency, the effective width of the |HPSF| curve, denoted by w(fsp) corresponds to the width of the main lobe, that is the distance between the two zero crossings on either side of the main peak as shown in
w(fsp)=2.44F#fsp−1, (5)
where fsp denotes spatial frequency in cycles per mm.
Now referring to
A measure of image brightness Bcolor(n3(k)) is also extracted for normalizing the sharpness measure to changes in image brightness. The focus value or sharpness measure considered here is the average of the output energies of the horizontal and vertical band-pass filters normalized by the square of the brightness value. Note that other sharpness measures can be used within the framework introduced here.
Due to the use of a Bayer-pattern image, the quincunx sampled green plane from yz(n(k)) is structurally separated into rectangular GR and GB color planes and processed as shown in
FBPF(xj(k)) can be viewed as a surface on the 2D focus actuator-spatial frequency plane, and is thus referred to here as the focus value surface. The FS AF Search Algorithm 406 takes in a set of FS AF Search Parameters 408 that includes M band-pass center spatial frequencies, fc,i, and discrete step-size magnitudes, Δn3,i, derived from the camera or device parameters 410 stored in memory 108. The device parameters 410 may include F#, the F-number of the lens, T1,2, the pixel sampling period of the image sensor in millimeters (mm) per pixel, and T3, the focus actuator sampling period in mm per focus actuator step. Based on the above, the processor 112 can automatically determine the search parameters via the effective width relationship of equation (5).
Referring now to
Now referring now to
The mathematical formulas underlying the method 600 will now be described in detail:
Step 1 (502): Periodically sample the digital spatial frequency axis in the interval T4≦fsp′≦½ with a spacing of T4 cycles per pixel between successive samples to produce a set of spatial frequencies
fsp′=n4T4, (7)
for n4 varying from 1 to N4−1, where N4=(2T4)−1+1. Uniform, rectangular sampling of the focus actuator-spatial frequency plane can be achieved by selecting T4=T3.
Step 2 (602): Convert the set of spatial frequencies fsp′ in cycles per pixel to those in cycles per mm via the pixel sampling period, T1,2, according to
Step 3 (504): The effective width in focus actuator steps for a given spatial frequency fsp′ can be derived by substituting equation (8) into equation (6) and dividing by the focus actuator sampling period T3, that is
Step 4 (604): For a fixed n4, periodically sample the effective width with a spacing of
steps in between successive samples, where T5 is an adjustable parameter to create a balance between the AF search speed and accuracy. The spacing in equation (10) is essentially a step-size associated with a focus value curve extracted from the spatial frequency of equation (7).
Step 5 (606): Since focus actuator steps are positive integers, quantize equation (9) to produce
Δwq(n4)=Q[Δw(n4)], (11)
where Q[.] is the round-to-nearest integer operator. If for any n4, equation (11) is 0, its value is replaced with the next minimum value of equation (11) which is greater than 0.
Step 6 (506): Use the digital curve of equation (11) to segment the spatial frequency axis into a set of E4 candidate band edges, fedge(e4), with e4 indicating band edge index, using the following first-difference operator ∂n
∂n
for n4 varying from 1 to N4−2. The band edge frequencies are then selected as follows using equation (12):
(i) fedge(0)=fmin, (13)
Step 7 (608): To avoid over-segmentation of the spatial frequency axis, merge the candidate band edges to eliminate narrow bands as follows:
If fedge(E4−1)−fedge(e4)>2T4, then the e4th edge is kept, otherwise it is removed. The last edge is always kept in the list, but to prevent over-segmentation of the highest frequency band, this check is done fedge(E4−1)−fedge(E4−2)<4.5T4, and if true, the edge E4−2 is also removed.
After this process is completed, E4 is set to the total number of remaining edges.
Step 8 (508): From the set of E4 band edges, construct a set of M=E4−1 filters or a filterbank with the following lower and upper band edges and center frequencies
fl,i=fedge(i) (17)
fu,i=fedge(i+1) (18)
and
fc,i=½(fu,i−fl,i), (19)
for i ranging from 0 to M−1.
Step 9 (510): Finally, compute a set of step-sizes associated with the ith band by substituting the center frequencies of equation (19) converted into units of cycles per mm via division by T1,2 in equation (6) as indicated below
for i ranging from 0 to M−2, and the minimum step-size of
which is defined with the consideration of the maximum tolerable blur circle diameter εCoC and ┌.┐ denoting the ceiling operator. In general, if the blur circle spreads over two pixels, the image information is lost. For a Bayer pattern sensor, this translates into the blur circle spreading over two 2×2 sets of Bayer pixels. Thus, εCoC is set to four times the diagonal of a square pixel, that is εCoC=4T1,2√{square root over (2)}.
The above segmentation procedure answers the questions of how many digital band-pass filters are necessary for a passive AF system and what step-sizes should be used for a focus value curve extracted from a given band-pass filter. In order to realize the derived M-band filterbank, for example fourth-order IIR Butterworth digital band-pass filters can be designed based on the set of lower and upper band-edge spatial frequencies. Such filters can be implemented efficiently using a cascade of second-order, direct-form II filter sections [18]. It is known that the general form of the spatial impulse response of the 1D horizontal (similarly for vertical direction) fourth-order IIR Butterworth digital band-pass filter can be written as the sum of two digital cosine waveforms as follows [18]:
where parameters α, β, γ, τ1, τ2, θ1, and θ2 are defined in terms of the filter's two pairs of complex conjugate poles, which are in-turn functions of the band-edge frequencies of equation (17) and equation (18).
Referring now to
Given the M-band filterbank defined in terms of the center spatial frequencies in equation (19) and the step-sizes in equation (20) and equation (21), a sharpness search algorithm is formulated to switch amongst the set of filters, two at a time, and decrease the step size automatically as the search gets closer to the peak location by utilizing local estimates of the first and second derivatives of the focus value surface FBPF(xj(k)), without the use of thresholds or difficult-to-set parameters. Although there are M bands, it should be pointed out that the focus values of all M bands need not be computed simultaneously. The search is formulated to utilize only two bands out of M at any given moment loaded into filter computation slots 0 and 1, thus reducing the amount of computational complexity and keeping the search compatible with existing passive AF hardware solutions, which usually incorporate two filters whose outputs can be computed simultaneously from the incoming stream of Bayer pattern image data. Note that more than two bands can be computed simultaneously if the device has more than two digital band-pass filters. As illustrated in
Now referring to
Referring now to
If a switch condition is not satisfied, as determined in decision block 1018, the process proceeds to the next iteration (k=k+1) in block 1020 and repeats blocks 1006-1018 as previously described. If the switch condition is satisfied, as determined in decision block 1018, the switch matrix to be used in the next iteration is computed in block 1022 and the analysis filters to be used at the next iteration are computed in block 1024. The sample state matrix is updated in block 1026 if the switch condition is satisfied. The step size is updated in block 1028 and the focus actuator position is updated in block 1030. Thereafter, the process loops back to decision block 1006 to determine whether the stop condition is satisfied. This method can be implemented by a computer readable medium encoded with a computer program wherein the steps are performed by one or more code segments.
The mathematical formulas underlying the method 1000 will now be described in detail. Considering the kth iteration, the logic behind the Filter-Switching Search algorithm can be compactly formulated using matrix-vector notation as follows:
Step 1 (1006): Check Stop Condition II (stopping conditions are described later in the section).
Step 2 (1010): Update the sample state matrix Φ(k) with the kth focus value samples as
Φ(k)=Φ(k−1)L3+[0[FBPF(x1(k)) FBPF(x0(k))]T] (23)
where Φ(k) is a 2×3 matrix of samples from the focus value surface, L3 is a 3×3 left-shift matrix with ones on the lower diagonal, 0 is a 2×2 matrix of zeros, FBPF(x0(k)) and FBPF(x1(k)) are the latest focus values computed at step-position n3(k) using the filter associated with the center spatial frequencies f0(k) and f1(k). The constructed state matrix Φ(k) is a sliding window which keeps track of the latest six samples of the focus value surface for differential analysis.
Step 3 (1012): Normalize the samples in the state matrix Φ(k) for comparison of the two focus values via
Φ′(k)=diag((FBPF(x1(k)))−1,(FBPF(x0(k)))−1)Φ(k), (24)
where diag(.) indicates a 2×2 diagonal matrix operator.
Step 4 (1014): Check Stop Condition I.
Step 5 (1018): Compute local estimates of the second partial derivatives of the focus value surface expressed as
are the estimates of the second partial derivatives, and sjT is a selection vector defined in Step 10, which is used to switch amongst the values from filter slot 0 and 1.
Step 6 (1020): Compute the switch condition given by
where switch kswitch denotes the iteration at which the switch condition is satisfied. The switch condition is the heart of the FS search and consists of two criteria based on analyzing the local estimates of the second differential behavior of the focus value surface.
The first condition,
is a check for negative curvature along the focus actuator movement direction. If the curvature is negative, it means that the search has moved to a position past the inflection point (or the curve of zero Gaussian curvature) on the current slice of the surface and is nearing the peak for the focus value curve based on the spatial frequency band under analysis. In order to slow down the search, the spatial frequency band is increased and the step-size decreased to allow the search to reduce the amount of overshooting of the peak as it nears the in-focus position.
The second condition,
is analogous to comparing estimates of the slopes of the two focus value curves under analysis. If the estimated slope of the focus value curve from the higher spatial frequency filter f1(k) is steeper than that of the focus value curve of the lower spatial frequency filter f0(k), it means that the search is getting closer to the peak. The search is slowed down by switching to a focus value curve for a higher spatial frequency band and its corresponding smaller step-size.
Step 7 (1022): Compute the switch matrix S(k+1) for the next iteration as indicated below
where S(k+1) is a 2×M matrix which is initialized to
at the beginning of the search and UM is an M×M upper-shift matrix with ones on the superdiagonal. If the filter loaded into slot 0, fo(k) is equal to fc,M−2, then there is no need to switch to another two sets of filters as the algorithm has converged to the last two filters in the filterbank.
Step 8 (1024): Compute the analysis filters to be used at the next iteration as
f(k=1)=S(k+1)fc (31)
where fc=[fc,0 fc,0 . . . fc,M−1]T is the vector of M spatial center frequencies, and f(k) denotes the vector whose elements are the center spatial frequencies for the two filters loaded into filter slots 0 and 1: f(k)=[f1(k) f0(k)]T.
Step 9 (1026): Update sample state matrix if the switch condition is satisfied as indicated below
where L2 is a 2×2 down-shift operator with ones on the lower diagonal. This operation incorporates the loading of the new filter into slot 1 and computing its focus value sample at the current focus actuator position. The existing set of three samples from the previous slot 1 filter are moved down one row (removing the samples from the previous slot 0 filter), while the new sample for the new filter is added to the state matrix.
Step 10 (1028): Update the step-size as follows:
where Δn3 denotes the vector of M step-sizes, Δn3=[Δn3,0 Δn3,1 . . . Δn3,M−1]T and the vector sjT denotes the filter slot selection vector where
is the index of the active filter slot, and s0=[1 0]T, s1=[0 1]T. When the switch condition is satisfied, the step-size of the new filter is loaded for the next iteration. If the step-size has not converged to the smallest step-size, then the step-size is incremented by one step provided that the switch condition is not satisfied. Once the step-size has converged to the smallest size, it is not incremented any further. The selection vector sjT is used to switch from filter slot 0 to filter slot 1 when the algorithm has converged to the last two filters in the filterbank.
Step 11 (1030): Update the focus actuator position to
n3(k+1)=n3(k)+Δn3(k+1) (35)
Step 12 (1006): Check Stop Condition II.
The stopping condition for the FS search is another key aspect of the algorithm and consists of two stop condition checks described next.
Stop Condition I: If Φ′(k) is fully populated, then compute an estimate of the following first partial derivative along the focus actuator direction for the filter slot determined by sjT,
where {circumflex over (d)}I(k) is an estimate of the first derivative formed by taking the average of two backward finite differences as given by equation (37):
If there is an indication of a slope sign change as given by
then the following procedure is performed to determine whether to continue or end the search by validating the sign change. The iteration at which the maximum focus value occurred for the currently loaded filter is first obtained from the following optimization
where k′switch denotes the iteration in which filter fj(k) was first loaded. The following conditions are then checked:
Stop Condition II: This check is done when the search has reached the end of the search range or if the search continues to move to a position beyond the end of the search range. First, equation (38) is used to determine the iteration at which the maximum focus value occurred for the currently loaded filter fj(k). The following conditions are then checked to determine the final in-focus position, {circumflex over (n)}3(k*):
Due to the use of the non-uniform step-sizes of equation (33) during the search, the following general form for inverse quadratic interpolation can be used to determine the final in-focus position using one vector of samples ρ, and two vectors of weights, ξ1 and ξ2,
The index kmax denotes the iteration at which the maximum sample occurred for the filter slot's samples that are being subjected to interpolation processing.
To obtain a better understanding of the Filter-Switching Search algorithm, a schematic view is provided in
The convergence comparison of the FS search with the Global Search (GS), Rulebased Search (RS), Modified RS (MRS), and the Switching Search (FS) in accordance with one embodiment of the present invention is illustrated in
The performance of any passive AF search in accordance with the present invention can be characterized by the following four metrics namely, (i) Total Number of Iterations (TNI), (ii) Total Distance Moved (TDM), (iii) Offset From Truth (OFT), and (iv) In-focus Position Overrun (IPO), which are defined as
where n3(k*) denotes the true in-focus position (normally determined by the GS method), n3(K−1) denotes the final estimated in-focus position, and n3(K−2) is the position corresponding to one iteration before moving to the final estimated in-focus position.
TNI gives the amount of iterations required for the search to converge and move to the estimated in-focus position. Thus, it is a measure of AF search speed. TDM gives the amount of steps traveled during the entire search from start to end. This measure can be regarded as a measure of power consumption, noting that larger distances moved require more power consumption by the focus actuator. OFT refers to the residual error of the search algorithm in bringing the object of interest into focus which is defined as the absolute difference between the estimated in-focus position and the GS-found in-focus position. Thus, it is a measure of AF search accuracy. The offset is characterized in terms of the maximum allowable offset of
steps defined by one-half the depth of focus. If the offset is less than the maximum offset, the AF search can be considered to have a high level of in-focus position accuracy.
Finally, IPO provides the amount of steps the search passed the true in-focus position before realizing the peak was passed and then returning to its estimated in-focus position location. This is a measure of user discomfort for a passive AF system, as consumers tend to complain if the passive AF search passes the peak by a noticeable amount before going back to the estimated in-focus position.
The performance of passive AF systems diminishes in low-light and low object contrast levels. Hence, it is important to gauge the robustness of an AF system across all combinations of different lights (1.5K, 640, 300, 160 lux) and object contrast levels (high 100%, mid 75%, low 50%, very low 25%). Since target object contrast is not known a priori, AF performance can be assessed by averaging the above performance measures across these contrast levels. The GS-found in-focus position from the combination of the highest contrast and light level is to be taken as the true in-focus position.
Without limiting the scope of the present invention, various experiments were performed to test various embodiments of the present invention. For example, a controlled experimental setup is recommended as a testing environment such as a darkroom which helps to shield out any unwanted lights and reflections. In the darkroom, custom ISO 12233 test charts of different target object contrast levels [19] can be fixed to a wall at a certain distance from the camera. The chart is then illuminated with two lights on either side of the camera whose lux level can be adjusted using digital control. To allow flexibility in testing, it is recommended to record the draft-preview raw Bayer-pattern sensor data at each focus actuator position, a Bayer-pattern focus sequence, for any AF search simulation analysis.
In order to gauge the effectiveness of the Filter-Switching AF framework, the AF test methodology of the present invention was applied to three different cameras: Camera 1, 2, and 3. Cameras 1 and 2 were prototype DSC type cameras with stepper motor focus actuators while Camera 3 was a prototype cell-phone type camera with a voice-coil focus actuator. Camera 1 had a 3 MPix CCD image sensor paired with a 3× optical zoom lens, while Camera 2 had a 10 MPix CCD image sensor also paired with a 3× optical zoom lens. Camera 3 had a 5 MPix CMOS image sensor and did not have an optical zoom lens. The relevant camera specifications required by the Filter-Switching AF framework are listed in Table I and can be easily determined from the data sheets for any given camera.
From the specifications in Table I, the sets of M band-edge frequencies in units of cycles per pixel and step-sizes in units of focus actuator steps were derived using the methodology of Section 2 with fmin set to 0.0250 cycles per pixel. In order to determine the best T5 parameter for use in equation (10), an initial test was performed by varying T5. In general, as T5 decreases, M increases and the resulting step-sizes become larger, leading to a faster but less accurate search with larger position overruns. As T5 increases, M decreases as well as the step-sizes, leading to a slower but more accurate search with less position overruns. By examining the 2D joint histogram of OFT and TNI, it was found that T5=6 provided a balance between accuracy and speed across all the camera platforms. The derived M-band filterbank with T5=6 was thus realized using the fourth-order IIR Butterworth digital band-pass filters and used to compute the focus value surface for the Filter-Switching algorithm.
In applying the test methodology of Section 4, the performance of Camera 1 and 2 was examined in both their respective Wide and Tele optical zooms. Camera 1 and 2 were placed at a distance of 1.5 meters from the wall-mounted contrast test chart, while due to the limitations of Camera 3, it was placed within <50 cm from the chart. Draft-preview, raw Bayer-pattern focus sequences were recorded for all the test conditions and then were analyzed to compute the average of the AF performance measures for the GS, RS, MRS, and FS searches as shown in Tables II, III, and IV. Note that only the results for Wide zoom are shown here, although a similar trend was found for Tele zoom. In the case of GS, RS, and MRS, the first difference filter was used to generate the focus value curve during the search as recommended in [3]. Also of note is that in the case of Camera 3, the test focus sequences were taken at every 10 focus actuator steps, reducing its T3 to 0.01 mm/step.
From the tables, it is evident that the FS search consistently outperformed all the other methods by reducing the total number of iterations, total distance moved, and position overrun while maintaining a high-level of accuracy. Note that OFT for GS is not necessarily zero since the true in-focus position is taken to be the GS position for the combination of the highest contrast and highest lux, and thus can vary for lower contrast and lower lux levels. It is worth emphasizing that the Filter-Switching framework automatically balances the AF speed versus accuracy tradeoff without tedious tuning of hard-to-set thresholds or parameters. While all methods produce offsets below the maximum tolerable offset, FS is able to significantly speed up AF while maintaining a high level of accuracy by taking advantage of the relationship between focus actuator step-size and digital spatial frequency and the differential properties of the resulting focus value surface during the search. A visual comparison of the speed and accuracy of the developed FS search can be seen in
In addition, in order to test the validity of the derived M band filterbank of Section 2, the performance of the FS search using the derived filterbank was compared with that of using a standard octave-band decomposition by varying M from 3 to 8 and computing the fourth-order IIR Butterworth digital bandpass filterbanks with the following band-edge frequencies for i=0 to M−1: fl,i=2(i−1)Δ
The step-sizes used with each of the filters in the bank were determined using their center frequencies via equation (20). The examination of the 2D joint histograms of OFT and TNI revealed that the FS search algorithm with the devised filterbank achieved the lowest offset and lowest total number of iterations in comparison to arbitrarily derived octave-band filterbanks. In essence, the present invention provides an optimum M-band filterbank which best captures the relevant spatial frequencies for a given set of camera specifications. In other words, there is no need to perform tedious experimentation to determine the best set of filters to use for a given camera as was previously recommended [20].
In another example, the results from a real-time search performance comparison for three prototype digital cameras are shown below:
The FS method outperforms the GS and MRS methods in every relevant performance metric. Moreover, the FS method should provide significant gain in convergence time on any camera platform.
As shown above, the performance results across various cameras exhibit the effectiveness of the present invention. The present invention provides a method of automatically setting the parameters of the passive auto-focusing system used in digital and cell-phone cameras. An auto-focus system model has been presented leading to the design of a set of optimal digital band-pass filters paired with a set of step-size magnitudes. Based on the devised model, the present invention provides a new Filter-Switching Search algorithm which provides a balance between speed and accuracy without the need for carrying out tuning or adjustment of hard-to-set parameters.
It will be understood by those of skill in the art that information and signals may be represented using any of a variety of different technologies and techniques (e.g., data, instructions, commands, information, signals, bits, symbols, and chips may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof). Likewise, the various illustrative logical blocks, modules, circuits, and algorithm steps described herein may be implemented as electronic hardware, computer software, or combinations of both, depending on the application and functionality. Moreover, the various logical blocks, modules, and circuits described herein may be implemented or performed with a general purpose processor (e.g., microprocessor, conventional processor, controller, microcontroller, state machine or combination of computing devices), a digital signal processor (“DSP”), an application specific integrated circuit (“ASIC”), a field programmable gate array (“FPGA”) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. Similarly, steps of a method or process described herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. Although preferred embodiments of the present invention have been described in detail, it will be understood by those skilled in the art that various modifications can be made therein without departing from the spirit and scope of the invention as set forth in the appended claims.
This patent application is non-provisional patent application of U.S. provisional patent application Ser. No. 61/148,042 filed on Jan. 28, 2009 and entitled, “Automatic Focusing Apparatus and Method for Digital Images Using Automatic Filter Switching” which is hereby incorporated by reference in its entirety.
Number | Name | Date | Kind |
---|---|---|---|
4990947 | Komiya et al. | Feb 1991 | A |
5115262 | Komiya | May 1992 | A |
5144450 | Kikuchi et al. | Sep 1992 | A |
5231443 | Subbarao | Jul 1993 | A |
5260736 | Toji | Nov 1993 | A |
5475429 | Kodama | Dec 1995 | A |
5689740 | Uchiyama | Nov 1997 | A |
5715483 | Omata et al. | Feb 1998 | A |
6701075 | Ogino | Mar 2004 | B2 |
6970646 | Hirai | Nov 2005 | B2 |
7079763 | Hirai | Jul 2006 | B2 |
7085488 | Ogino | Aug 2006 | B2 |
7110035 | Hellstrand | Sep 2006 | B1 |
7173243 | Yang et al. | Feb 2007 | B1 |
7236696 | Higuma et al. | Jun 2007 | B2 |
7389042 | Lin et al. | Jun 2008 | B2 |
7783098 | Douglass et al. | Aug 2010 | B2 |
20070268376 | Yoshikawa et al. | Nov 2007 | A1 |
20100045825 | Hatori et al. | Feb 2010 | A1 |
Number | Date | Country | |
---|---|---|---|
20100188558 A1 | Jul 2010 | US |
Number | Date | Country | |
---|---|---|---|
61148042 | Jan 2009 | US |