Method of image processing

[0002] The present invention relates to methods of image processing, and more particularly to methods for removing the degradation resulting from sampling systems where more than one channel of information is captured and both channel signal and sensitivity are unknown and vary at each sample location.

[0003] An example of a system that captures more than one channel with unknown and varying signal and sensitivity is a color filter array (CFA) film that has been spatially sampled with a film scanner. A CFA film for capturing the red, green, and blue channel information required for human color perception contains at least three colored species from which red, green, and blue channel information can be decoded. A CFA film can be designed for direct viewing and/or optical printing. However the CFA structure degrades image quality.

[0004] A CFA film is disclosed in EP 935168. The color filter array comprises colored resin particles which are randomly arranged.

[0005] If the number of independent channels of information acquired by electro-optical scanning of a CFA film is at least twice minus one the number of desired information channels, then at each scanned pixel location, it is possible to determine both the amount of signal (silver or other light modulating species formed in a photo process as a function of light intensity) formed in spatial coincidence with the total sub-pixel area occupied with similarly colored CFA elements (the relative sensitivity). The twice minus one arises from the area constraint at each pixel; that is that the sum of the area of all of the individual sub-pixel CFA elements is equal to the effective scanning aperture or pixel area. One such system is described in U.S. patent application Ser. No. 09/080,791 wherein twice the number of desired channels of independent information is acquired in order to essentially (within the limits of bit depth) determine both the amount of silver formed under the total sub-pixel area filled with similarly colored CFA elements and the total sub-pixel area filled with similarly colored CFA elements.

[0006] However, most existing scanning systems do not possess the capability of acquiring these extra channels of independent information required to perform this joint determination. By applying the following image processing method, it is possible to determine the amount of silver formed under the total sub-pixel area filled with similarly colored CFA elements and the total sub-pixel area filled with similarly colored CFA elements from electro-optical scanning where the number of independent channels of information acquired equals the number of desired information channels.

[0007] For a three channel system the problem to be solved, at each pixel position (p, q),

[0008] can be expressed as follows:

[0009] C_m1(p, q)=X₁(p, q)C₁(p, q)  (1a)

[0010] C_m2(p, q)=X₂(p, q)C₂(p, q)  (1b)

[0011] C_m3(p, q)=X₃(p, q)C₃(p, q)  (1c)

[0012] X₁(p, q)+X₂(p, q)+X₃(p, q)=1.0  (1d)

[0013] where C_miis the measured transmittance of the ith color channel, X_i

[0014] is the fraction of effective scanning pixel area covered by sub-pixel CFA

[0015] elements of the ith color, and C_iis the desired recorded image signal for the ith channel at each pixel position (p, q). For simplification, the (p, q) will not be included explicitly in subsequent equations, but it will be understood that all equations are a function of pixel position.

[0016] A simple assumption would be that for each channel i in a local area, the average fraction of pixel area covered, the values of X_i, are constant and equal to the average fraction observed over the entire scanned image and these fractions are known. Defining these known fractions as {overscore (X)}_i, and assuming that over some small area that essentially defines the window for a low pass filter, the values of C_iare calculated as follows:

[0017] C_i=1pf(C_mi)/{overscore (X)}₁  (2)

[0018] where 1pf(C_mi) is the low-pass version of the C_mimeasurement, sufficiently low-passed so that the above average fraction assumption is constant. Low-pass filtering of a sampled image is well known in the art and is achieved, for example, by applying a finite impulse response (FIR) filter to a sampled signal.

[0019] Unfortunately, in order to achieve this constancy condition, excessive low-pass filtering is required, resulting in low spatial resolution and image quality. Alternate assumptions can be employed that result in improved image quality.

[0020] Another approach well-practiced in the art of color imaging is to transform the three measured signals, e.g. red, green, and blue, to signals such as one luminance and two chrominance signals. Luminance is broadly defined as a linear combination of red, green, and blue, all with positive weightings that sum to unity. The two chrominance signals are broadly defined as independent representations of the color information with weightings that sum to zero. A set of weightings that is used in television systems is:

[0021] L=0.30C₁+0.59C₂+0.11C₃  (3a)

[0022] B−L=−0.30C₁−0.59C₂+0.89C₃  (3b)

[0023] R−L=0.70C₁−0.59C₂−0.11C₃  (3c)

[0024] where L is the luminance-like signal and B-L and R-L are the two chrominance-like signals. The weighting coefficients for L are all positive and sum to unity. The weighting coefficients for B-L and R-L sum to zero. For this example, C₁=R (red), C₂=G (green) and C₃=B (blue). Owing to the characteristics of the human visual system, it is possible to low-pass filter chrominance signals significantly more than luminance signals whilst maintaining good image quality. Applying this transformation to the unknown signal, unknown sensitivity system described by equations (1) produces an improved image, i.e. an image without high frequency colored noise. The noise owing to the color filter array is reduced in some colored areas. However, most of the image suffers from CFA noise and colored mottle (low frequency color noise). The high frequency CFA noise is manifest as non-colored noise owing to the transformation, equation (3a), and lack of low-pass filtering of the L signal. If the CFA has low frequency power, then the resulting chrominance signals will contribute colored mottle (low frequency noise) to an image reconstructed after the above described image processing. If all of the transformed signals in equations (3) were equally low-pass filtered, then the result, by the application of linear systems theory, would be equivalent to that achieved by low-pass filtering the signals represented by equation (2).

[0025] It is an aim of the invention to provide a method of image processing which avoids the problems mentioned above, which does not require a special scanner but gives good quality results with typical scanning means.

[0026] According to the present invention there is provided a method of forming output signals from a plurality of channel input signals, said input signals being a function of both unknown recorded signal levels and unknown signal sensitivities, the method comprising the steps of;

[0027] sampling and measuring the product of the recorded signals and the signal sensitivities for each sampled channel input signal;

[0028] determining at each sampled location channel weighting values which eliminate the contribution of the unknown signal sensitivities; and

[0029] further processing the plurality of channel input signals using the determined channel weighting values to form output signals that represent the unknown recorded signal levels.

[0030] By applying multi channel scanning and image processing according to the invention it is possible to remove essentially all of the image quality degradation due to the CFA structure.

[0031] FIG. 1 is a flow diagram showing the method of forming a pseudo-luma signal; and

[0032] FIG. 2 is a flow diagram showing the method of forming color signals.

[0033] The method of this invention starts with a first step of producing a pseudo-luma signal, see FIG. 1, and then follows with steps to form the color signals, see FIG. 2.

[0034] In general terms, with reference to FIG. 1, the image is scanned, step S1, and samples taken of each channel. Each sample is spatially transformed, step S2. Step S2 is not an essential step and in some methods of the invention this step is omitted. Depending on the method of the invention being used a high pass or a low pass filter may be used. Weighting values are determined in step S3. It is important to note that the neighborhood used to determine the weighting values need not be the same neighborhood as that used in the spatial transformation. In method A, described in detail later, the neighborhood used for determining the weighting values covers a single pixel. In step S4 a pseudo luma equation is formed using samples taken from each channel and the weighting values determined in step S3.

[0035] A pseudo-luma signal is defined as any sum of C_i(i=1 to 3) signals that satisfies the following equation:

[0036] Pseudo-Luma=k₁C₁+k₂C₂+k₃C₃  (4)

[0037] where the k values, i.e. the channel weightings, are constrained to sum to unity:

[0038] k₁+k₂+k₃=1.0  (5)

[0039] If the image signal color is constant in a small neighborhood (small number of pixels) which is oftentimes the case and is also the case with the human visual system, then it can be shown that the noise owing to the CFA can be eliminated if the appropriate values of k_iare chosen for each pixel location. Constant color, CC_i, is defined to be the fraction of C_i, i.e.

[0040] CC_i=C_i/(C₁+C₂+C₃  (6)

[0041] where C_iis the transmittance. If constant color is maintained in a region, then the values of k_ican be adjusted so that the calculated Pseudo-Luma is independent of the X_i's in that region. An additional constraint on the values of k_iprovides this:

[0042] k₁C₁=k₂C₂=k₃C₃  (7)

[0043] When the values of k_iare adjusted to achieve the above dual equality (equation (7)), then the contribution of the unknown signal sensitivity proportion or the X_i's is eliminated in that region of assumed constant color. By substitution of the above equation (7) with equations (1) and (4):

[0044] k₁C_m1=X₁k₁C₁  (8a)

[0045] k₂C_m2=X₂k₂C₂=X₂k₁C₁  (8b)

[0046] k₃C_m3=X₃k₃C₃=X₃k₁C₁  (8c)

[0047] Adding the equations (8a-c) yields:

[0048] k₁C_m1+k₂C_m2+k₃C_m3=(X₁+X₂+X₃)k₁C₁=K₂C₂=K₃C₃  (9a)

[0049] or from combining with equation (4):

[0050] Pseudo-Luma=3(k₁C_m1+k₂C_m2+k₃C_m3)  (9b)

[0051] where X₁+X₂+X₃=unity. If these values of k_ican be found, then the unknown signals (C₁) and Pseudo-Luma can be fully determined from the measured transmittance values C_mi.

[0052] There are various methods for determining these values of k_i.

[0053] Determining k_i—Method A

[0054] A first method of calculating the values of k_iis to assume a constant color determined from a low-pass filtered version of the measured transmittance values, C_mi. This is different from the above cited prior art example where the desired signals, C_i, are determined by simple low-pass filtering of the measured transmittance values, C_mi. Instead, the inverse of the low-pass filtered measured transmittances are used to form the weightings k_iwhich are then used to form the Pseudo-Luma and ultimately a colored image as will be described later on. The values of k_iare calculated as follows:

[0055] k₁=1pf(C_m2)1pf(C_m3)/[1pf(C_m1)1pf(C_m2)+1pf(C_m1)1pf(C_m3)+1pf(C_m2)1pf(C_m3)]  (10a)

[0056] k₂=1pf(C_m1)1pf(C_m3)/[1pf(C_m1)1pf(C_m2)+1pf(C_m1)1pf(C_m3)+1pf(C_m2)1pf(C_m3)]  (10b)

[0057] k₃=1pf(C_m1)1pf(C_m2)/[1pf(C_m1)1pf(C_m2)+1pf(C_m1)1pf(C_m3)+1pf(C_m2)1pf(C_m3)]  (10c)

[0058] where 1pf(C_mi) is a low-pass filtered version of the measured transmittance C_mi. The Pseudo-Luma's can be calculated by using the values of k_ifrom equations (10a-c) with equations (9) and (4). This method provides a Pseudo Luma signal with minimal noise contribution from the CFA. Whilst the 1pf(C_mi) values carry low frequency variations from the CFA, the effect of low frequency variations in the k_ivalues has a much smaller effect than low frequency variations in the 1pf(C_mi) values, owing to the normalizing and averaging properties of equations (10a-c). Furthermore as the k_ivalues sum to unity, see equation (5), the high frequency content and overall dynamic range of the Pseudo-Luma signal is not degraded.

[0059] Determining k_i—Method B

[0060] An alternative method for calculating the values of k_iis to assume that within a spatial neighborhood the ratios between high-passed measured signals are constant and any deviation from this constancy is owing to the CFA modulation. This assumption is realized by modifying the values of k_iso that the high frequency power in the measured transmittance values, C_miis minimised. Firstly, high-pass signals, HP_i, are formed:

[0061] HP₁=hpf(C_m1)  (11a)

[0062] HP₂=hpf(C_m2)  (11b)

[0063] HP₃=hpf(C_m3)  (11c)

[0064] by application of an FIR or other digital filtering means.

[0065] High spatial frequency power, HFP is defined as:

[0066] HFP=Σ[k₁HP₁+k₂HP₂+k₃HP₃]**2  (12)

[0067] where Σ denotes summing in a weighted neighborhood about the sample for which the values of k_iare being determined. Equation (12) is minimised by taking partial derivatives with respect to the values of k_iand solving the resulting simultaneous equations. By simple inspection of equation (12), if all of the k_ivalues equal zero, then HFP is minimised. However this would violate equation (5) which requires that the k_ivalues sum to unity. Using equations (5) and (12) to eliminate k₃yields:

[0068] HFP=Σ[k₁(HP₁−HP₃)+k₂(HP₂−HP₃)+HP₃]**2  (13)

[0069] Equation (13) is minimised by taking partial derivatives with respect to k₁and k₂and solving the following equation for k₁and k₂. $\begin{array}{cc}\left[\begin{array}{cc}\sum {\left({\mathrm{HP}}_1-{\mathrm{HP}}_3\right)}^2& \sum \left({\mathrm{HP}}_1-{\mathrm{HP}}_3\right)\left({\mathrm{HP}}_2-{\mathrm{HP}}_3\right)\\ \sum \left({\mathrm{HP}}_1-{\mathrm{HP}}_3\right)\left({\mathrm{HP}}_2-{\mathrm{HP}}_3\right)& \sum {\left({\mathrm{HP}}_2-{\mathrm{HP}}_3\right)}^2\end{array}\right][\begin{array}{c}{k}_1\\ k_2\end{array}]=\left[\begin{array}{c}-\sum \left({\mathrm{HP}}_1-{\mathrm{HP}}_3\right){\mathrm{HP}}_3\\ -\sum \left({\mathrm{HP}}_2-{\mathrm{HP}}_3\right){\mathrm{HP}}_3\end{array}\right]& \left(14\right)\end{array}$

[0070] k₃is determined with known values of k₁and k₂, from equation (14), and from equation (5).

[0071] Equations similar to equation (13) can be formed by choosing to eliminate k₁or k₂instead of k₃. When applied, slightly different values of k_iare obtained. An extension of this method would be to form and average the 3 sets of k_ivalues formed when k₁, k₂, or k₃are eliminated by the constraint of equation (5).

[0072] Determining k_i—Method C

[0073] Yet another method for calculating the values of k_iis to assume that the standard deviations values, SD_i, in a neighborhood are made equal by weighting the measured signals by the k_ivalues subject to the constraint of equation (5):

[0074] SD₁=sqrt{(Σ[C_m1−C_m1(av)]**2)/(n−1)}  (15a)

[0075] SD₂=sqrt{(Σ[C_m2−C_m2(av)]**2)/(n−1)}  (15b)

[0076] SD₃=sqrt{(Σ[C_m3−C_m3(av)]**2)/(n−1)}  (15c)

[0077] where n is the number of pixels in the neighborhood used to evaluate the values of SD_i. C_mi(av) is the average value of C_miin this region. In the most general sense equations (15a-c) could be modified to unequally weight the contribution of the members of the neighborhood to the standard deviation calculation. The values of k_ithat equate standard deviation and satisfy equation (5) and equations (15a-c) are:

[0078] k₁=SD₂SD₃/(SD₁SD₂+SD₁SD₃+SD₁SD₂)  (16a)

[0079] k₂=SD₁SD₃/(SD₁SD₂+SD₁SD₃+SD₁SD₂)  (16b)

[0080] k₃=SD₁SD₂/(SD₁SD₂+SD₁SD₃+SD₁SD₂)  (16c)

[0081] An extension of this method is to calculate the standard deviations based on spatially filtered versions of the values of C_mi. For example, high-pass filtering of the C_mi(as described in Method B) could be using to form the values of SD_ias follows:

[0082] SD₁=sqrt{([HP₁−HP₁(av)]**2)/(n−1)}  (17a)

[0083] SD₂=sqrt{([HP₂−HP₂(av)]**2)/(n−1)}  (17b)

[0084] SD₃=sqrt{([HP₃−HP₃(av)]**2)/(n−1)}  (17c)

[0085] where HP_i(av) is the average value, typically zero, of HP_iin this region.

[0086] Those skilled in the art will appreciate that combinations of the above methods represent additional methods. Furthermore, it will be appreciated that the definitions of neighborhoods is broad and includes adaptive, varying and other dynamic neighborhoods. Furthermore it is not required that neighborhoods be the same in any of these combinations, nor does the spatial extent of any of the FIR filters used have to be the same as the spatial extent of a neighborhood.

[0087] Having determined the values of k_iby any one of the above described methods or combinations thereof, Pseudo-Luma can now be determined at each sample location by application of equation (9b). The reconstructed Pseudo-Luma signal provides a luminance-like signal, but saturated colors have lower values than would be the case had a true luminance signal (that would result from application of equation (3a)) been calculated.

[0088] Method for Estimating Unknown Signals

[0089] The pseudo-luma equation and the determined weighting values are then used to determine the unknown signals, see step 5 in FIG. 2. The weighting values input, in step S3a, into step 5 need not be the same weighting values as those used to determine the pseudo-luma equation in step S4, i.e. it is not essential to use the same weighting values to determine the pseudo-luma equation and the color signals.

[0090] Simple application of equation (7), that exploits the color constancy in a neighborhood assumption, can be used to determine the estimates of the values of C_i.

[0091] Pseudo-Luma=k₁C₁+k₂C₂+k₃C₃  (4)

[0092] k₁C₁=k₂C₂=k₃C₃  (7)

[0093] Pseudo-Luma=3(k₁C_m1+k₂C_m2+k₃C_m3)  (9b)

[0094] C₁=[Pseudo-Luma]/[3k₁]=C_m1+k₂C_m2/k₁+k₃C_m3/k₁  (18a)

[0095] C₂=[Pseudo-Luma]/[3k₂]=k₁C_m1/k₂+C_m2+k₃C_m3/k₂  (18b)

[0096] C₃=[Pseudo-Luma]/[3k₃]=k₁C_m1/k₃+k₂C_m2/k₃+C_m3  (18c)

[0097] It will be appreciated by those skilled in the art that the above described methods contains ratios which may be sensitive to noise in signal and/or calculation means. Various known methods, such as low-pass filtering, median filtering, range limiting/thresholding, etc. may be applied to any of the above calculations in order to reduce any impact from noise.

[0098] The method described can be used for any system with unknown signals and unknown signal sensitivity as described by equations (1). Whilst equations (2) are written for a three channel system, those skilled in the art will appreciate that the method can be used for any multi-channel system where the sum of unknown signal sensitivities is known at each, equation (1d), and the image is measured with a number of channels that is equal to half the sum of the number of unknown signals plus the number of unknown sensitivities, equations (1a-c).

[0099] By applying the multi-channel scanning and image processing taught in this specification, it is possible to remove essentially all of the image quality degradation owing to the CFA structure. The method described is applicable to CFA systems where the effective sampling aperture at the film plane is greater than the feature size of elements of the CFA.

[0100] The invention has been described in detail with reference to several preferred embodiments thereof. It will be understood by those skilled in the art that variations and modifications may be effected within the scope of the invention.