Digital Image Noise Suppression Method Using SVD Block Transform

The invention relates to block transform digital image processing methods for reducing noise in a digital image.

U.S. Pat. No. 4,553,165, issued Nov. 12, 1985 to Bayer discloses a block transform image processing method for removing noise (such as film grain noise) from a digital image, produced for example by scanning a photographic image. According to the method disclosed by Bayer, the digital image is divided into blocks. Each block is transformed, for example by Walsh Hadamard transformation, to yield blocks of transform coefficients. The transform coefficients are modified in a nonlinear manner to reduce the noise in the image block, and the blocks of modified coefficients are inversely transformed to yield the processed image having reduced noise. The effect of this processing is to reduce the appearance of noise in the processed image while avoiding the introduction of an artifact caused by other prior art noise reduction processes in which false "edges" would appear in smooth areas of the the processed image (such as facial features).

While the technique disclosed by Bayer is indeed effective to avoid the appearance of false edges, while reducing the appearance of noise in the processed image, extensive testing by the present inventors of the image processing method employing the Walsh-Hadamard transformation has revealed the presence of another class of undesirable artifacts in the processed images that is not effectively avoided by the Walsh-Hadamard block transformation noise reduction method. These artifacts occur in areas of the image having fine texture, whereby the texture is replaced by smooth appearing areas. This artifact is particularly noticeable and objectionable in such image features as grass, hair, and textile patterns such as carpet. Careful investigation has also shown that the Bayer method somewhat reduces the sharpness of edge detail.

It is therefore the object of the present invention to provide a block transform image processing method that avoids the shortcomings noted above. In arriving at the present invention, we found it helpful to consider the statistical properties of the noise being removed from an image, and the statistical properties of the image details such as texure and edges that were to be preserved in the processed image. In particular, we examined the statistical properties of the noise and image detail in the transformed coordinate space. For a spatial transformation of the Walsh-Hadamard type, a transform coefficient of the noise is characterized by a generally Gaussian distribution around a mean value of zero. This is shown by Curve 10 in FIG. 2. The transform coefficients of the picture detail, including edges and texture, form a generally Lapacian distribution, also centered about zero (shown by Curve 12 in FIG. 2.)

The transform coefficients from the picture detail have generally higher amplitude in absolute terms than the ones From the noise. Noise suppression is achieved by thresholding the transform coefficients or by modifying them through a non linear gain function. This will remove most of the noise but unfortunately it will remove the low amplitude transform coefficients from the image detail which will create artifacts. The artifacts are most noticeable and objectionable in low contrast, fine textured area.

We also examined the variance distributions of the image components such as noise, texture, and edges, and noted that there was a much better separation of the statistics of the image components when plotted against variance of small regions. FIG. 3 is a graph showing variance plotted against distribution (number of occurrances) for film grain noise (Curve 1), texture (Curve 16), and edge detail (Curve 18) for a typical digital image produced by scanning a photograph.

It will be appreciated from a comparison of FIG. 2 with FIG. 3, that a noise reduction technique that discriminates based upon the variances of image detail will have a much better chance of reducing noise without affecting texture extensively. We also came to realize that there exists an image transformation called singular value decomposition (SVD) that decomposes an image into a set of singular vectors and singular values that are closely analagous to the concept of principal component analysis in statistics. This can be appreciated from the following analysis.

If an m.times.n matrix is treated as a set of n m dimensional column vectors and the mean column vector is set to zero by subtracting it from every column vector of the matrix, then the singular values of the resulting matrix are the square roots of the variances of the m vector components in a rotated space. The rotated space is such that there is no correlation between any two components of the sample vectors in the rotated space. The distribution of the singular values for noise is a slowly decreasing function when they are ordered in decreasing order. The distribution of the singular values for the picture detail will be quite different from the noise. Also, the singular values for the picture detail will be much higher than those of the noise. And as mentioned above, discrimination of the noise from the picture detail will be much better.

Armed with these insights we proceeded to develop a block transformation digital image processing method for reducing noise in an image employing singular value decomposition as the transformation. According to the method of the present invention, a digital image is processed in a computer to remove noise by performing the following steps. First a non linear gain function is produced based upon the measured statistics of the singular values of the noise in the image. A detail image and a low pass filtered image are produced from the digital image to be processed. The detail image is divided into blocks and the blocks are transformed into singular vectors and a diagonal array of singular values. The non linear gain function is applied to the singular values to produce an array of modified singular values. An inverse SVD transform is performed on the singular vectors and the modified singular values to produce blocks of processed detail image values. Finally, the processed detail image is added to the low pass image to generate a processed image having reduced noise.

FIG. 1 is a schematic diagram showing a digital image processing system suitable for practicing the method of the present invention;

FIG. 2 is a graph useful in describing the statistics of image features processed according to the prior art;

FIG. 3 is a graph useful in describing the statistical features of an image processed according to the present invention;

FIG. 4 is a block diagram illustrating the method of digital image processing according to the present invention;

FIG. 5 is a block diagram illustrating the step of generating the table of factors described in FIG. 4;

FIG. 6 is a graph showing the values of a typical table of factors generated according to the steps shown in FIG. 5;

FIG. 7 is a block diagram showing a block overlap method of processing a digital image according to the present invention;

FIG. 8 is a diagram useful for describing the image processing method shown in FIG. 7;

FIG. 9 is a block diagram showing a digital image processing method according to the present invention, having a plurality of stages for processing different spatial frequency pass bands of the image;

FIGS. 10a-10c are diagrams showing the values of the coefficients employed in the digital filters shown in FIG. 9;

FIG. 11 is a diagram showing a digital image processing method according to the present invention employing block overlap and a plurality of spatial frequency band pass image signals;

FIG. 12 is a diagram showing a mode of practicing the invention including means for processing diagonal edge information; and

FIG. 13 is a diagram useful in describing the operation of the digital image processing method shown in FIG. 12.

The digital image signal referred to in the following description is generated by scanning and sampling an original image. For purposes of describing the preferred embodiments, the input signal is generated from a photographic negative or transparency. The digital image signal represents a variety of spatial components of the image, including an average brightness level, fine detail such as lines and textures, intermediate detail such as small features, and coarse details such as shading on smooth surfaces and other gradually varying features. In addition, the signal includes a noise component affecting most of the spatial components of the image to some degree.

With a photographic negative or transparency, much of the noise is film grain noise. While the invention will be described in connection with sampled data from a photograph, it should be understood that the input signal can represent other information or data such as would be derived from directly scanning an object, from a composite video signal or from image information stored in optical or magnetic storage media. In such cases, the noise may originate in other characteristics of the image signal generating system. Since the singular values measure correlations, the method can remove noise originating from a wide variety of noise sources.

FIG. 1 is a schematic diagram showing a digital image processing facility useful for practicing the present invention. The digital image processing facility includes an input image scanner 20, such as a CCD scanner or a graphic arts flat bed or drum scanner. A digital image signal generated by an input scanner 20 is processed by the digital image processing computer 22. The digital image processing computer 22 can be a general purpose digital computer, or a special purpose computer specifically designed for processing images e.g. a parallel multi-processor computer with 16 micro-processors and local memory. The original digital image from the scanner, and/or the processed digital image may be stored in a mass image storage memory 24, comprising for example magnetic or optical disk storage media.

The original and/or processed image may be displayed by means of an output image scanner 26, such as a CRT or laser film scanner. The system is controlled by an operator from a work station 28 such as the Sun work station manufactured and sold by Sun Microsystems Inc. The work station includes a CRT 30 for temporarily displaying an image, a keyboard 32, and graphics input device such as a mouse and graphics tablet 34.

FIG. 4 is a block diagram showing the major steps implemented by the image processing computer 22 in carrying out one mode of digital image processing according to the present invention. A low pass digital filter (e.g. a 31.times.31 pixel Gaussian Filter) is applied (36) to the digital image signal Z to produce a low pass digital image signal G. The low pass digital image signal G is subtracted (38) from the suitably delayed digital image signal Z to produce a detail image signal H. The detail image signal H is processed, employing SVD transformation (40), as described in detail below to produce a noise reduced detail signal H'.

In the SVD process, the detail image signal H is block SVD transformed (44) employing the well known SVD computer program described on pages 229 to 235 of the book: Computer Methods for Mathematical Computations by G. E. Forsythe, M. A. Malcolm, and C. B. Moler published by Prentice Hall Inc., Englewood Cliffs, N.J., 1977, to produce singular vector matrices U, V.sup.T, and a diagonal matrix D of singular values d.sub.i arranged in order of descending amplitude where:

where:

H is an n.times.n sub block (e.g. 20.times.20 pixels) of the image,

U contains the eigenvectors of HH.sup.T,

D is a diagonal matrix which contains singular values d.sub.1 d.sub.2 . . . d.sub.20 in order of descending amplitude, and

V contains the eigenvectors of H.sup.T H.

The singular values in array D are modified in a nonlinear fashion (46) by factor f.sub.i stored in lookup table (LUT) 48 to produce an array D' of modifed singular values. The generation of the factors stored in the lookup table 48 will be described in more detail below. The array of modifided singular values D' and singular vectors U, V.sup.T are inversely transformed (50) to produce a noise reduced detail signal.

The generation of the factors f.sub.i from a nonlinear gain function (52 in FIG. 4) will now be described with reference to FIG. 5. A digital noise image N generated for example by scanning a uniformly exposed and developed film is low pass filtered (54) for example by a 31.times.31 pixel Gaussian digital filter, to produce a low pass filtered noise image L.

The low pass filtered noise image L is subtracted (56) from the suitably delayed noise image N to produce a noise detail image signal. The noise detail image signal is block SVD transformed (58) to produce singular vectors and arrays of singular values d.sub.i for the noise image blocks. The singular values d.sub.i from each block of the transformed noise detail image are accumulated and the means .mu..sub.i and standard deviations .sigma..sub.i of the singular values in the respective positions of the the array are calculated (60) as follows: ##EQU1## where i is an index for the order of singular values, and j is an index for different blocks.

A factor f.sub.i for each singular value d.sub.i is then generated (62) by considering the following facts. The singular values of the noise will be centered at .mu..sub.i with standard deviation .sigma..sub.i. The singular values of the shading areas of the image will have slightly higher values than those of the areas dominated by noise. The singular values of the textured areas of the image will have even higher values depending on the texture. The singular values of the edges in the image will have much higher values than those of the noise.

Considering the above, each singular value d.sub.i, is multiplied by a factor f.sub.i through a non linear function F(d.sub.i, .mu..sub.i, .sigma..sub.i) defined as below to produce the output singular value d': ##EQU2## The parameters a, p, and th1 are determined such that a good noise suppression in the shading and textured area is achieved The parameter th2 controls a threshold level when d.sub.i is large. The effect of the factor f.sub.i can be seen from the following discusions. When d.sub.i .ltoreq.(.mu..sub.i +th1*.sigma..sub.i). then the factor f.sub.i =0, and noise in the uniform area is suppressed. When d.sub.i >>(.mu..sub.i +th1*.sigma..sub.i), the factor f.sub.i approaches to 1-(th2*.sigma..sub.i /d.sub.i). When this factor is multiplied by d.sub.i, d'.sub.i =(d.sub.i -th2*.sigma..sub.i), resulting in some noise suppression in the edge region. The parameters a and p controls the curve shape for the transition between these two extremes The typical values for the parameters are: th1=3, th2=0 to 2,p=4,a=0.05.

The formula mentioned above is only one form of many possible non linear curve shapes that can be used as a factor for noise suppression. If one wants to remove noise partially, one could have the factor as follows: ##EQU3## where fmin is a constant chosen for the desired noise level reduction which has a value between zero and one. When fmin is equal to ##EQU4## the noise reduction provided by the present invention is equivalent to the improvement in film speed of 1 stop.

This function when applied to the singular values has an effect of thresholding the noise, reducing the singular values for shading area, and retaining textures and edges. The thresholding level and the curve shape depends on the artifacts one might tolerate.

The values f.sub.i of the nonlinear gain functions F(d.sub.i, .mu..sub.i, .sigma..sub.i) are calculated for each singular value d.sub.i to produce a table of factors f.sub.i for each singular value. The factors f are digitized in the form of look up tables and stored in look up table 48 shown in FIG. 4.

According to the presently preferred mode of practicing invention, the block SVD processing is performed using a moving average technique employing block overlap to reduce the appearance of blocking artifacts. FIG. 7 is a schematic block diagram showing the major steps involved in the block overlap SVD processing.

For the purpose of simplifing the description, processing incorporating a 4.times.4 pixel block, with a 2 pixel step in the horizontal and vertical directions will be described. Such a block overlap pattern is shown in FIG. 8. In the actual reduction of practice, a 20.times.20 pixel block was employed with 1 or more pixel steps, depending on the tolerance of the blocking artifacts. Referring to FIG. 7, an image detail signal H, generated as shown in FIG. 4, is processed by a block SVD process 40 (as shown in FIG. 4) to produce a processed image detail signal H.sub.1 '. Simultaneously, the image detail signal H is delayed by 2 pixels (64) and block SVD processed (40') to produce a processed image detail signal H.sub.2 '. The 2 pixel delay has the effect of shifting the blocks that are processed by 2 pixels, as shown by the blocks of pixels labeled 66 and 68 in FIG. 8. The image detail signal is similarly delayed by 2 lines (70), and 2 lines plus 2 pixels (72) and block SVD processed (40") and (40'") to produce processed image detail signals H.sub.3 ' and H.sub.4 ' respectively. The 2 line and 2 line plus 2 pixel delays have the effect of shifting the blocks as shown by the blocks of pixels labelled (74) and (76) respectively in FIG. 8. The processed detail signals H.sub.1 ', H.sub.2 ', H.sub.3 ' and H.sub.4 ' are registered, summed, and averaged (78) to produce the processed image detail signal H'. The processed image detail signal H' is added to the low pass filtered image signal G to produce the processed image signal Z' as shown in FIG. 4. It will be readily apparent that the processing method may be extended to larger blocks with different amounts of block overlap.

The SVD processing method according to the present invention can be extended to a multi-stage processing method of the type disclosed in U.S. Pat. No. 4,442,454 issued Apr. 10, 1984 to Powell, wherein each stage of the processing employs a detail signal representing a different pass band of spatial frequencies. The processed digital image signal is obtained by combining the processed detail signals from each stage, whereby noise from different spatial frequency content is effectively removed from the image. As shown in FIG. 9, and input digital image Z is filtered through a 3.times.3 pixel low pass filter 80 to obtain the low pass filtered image S. S is subtracted from Z (82) to obtain the difference image Z-S which is a bandpass filtered version of image Z. Similarly, the low pass filtered image S is filtered through a 5.times.5 pixel low pass filter 84 to obtain the low pass filtered image M, which is in turn filtered through a 9.times.9 pixel low pass filter 86 to obtain the low pass filtered image L. M is subtracted from S (88) to form the image difference signal S-M; and L is subtracted from M (90) to form the image difference signal M-L. The difference image signals Z-S, S-M, and M-L are all bandpass versions of the original image Z, with different spatial frequency contents. These bandpass images are processed by the SVD process 40, 40', and 40" respectively to reduce the noise in the different frequency bands before they are combined (92) with the low pass filtered image L to produce the output image Z'. The filter coefficients used to produce digital filters 80, 84, and 86 are shown in FIG. 10a, b, and c respectively. The 5.times.5 pixel low pass filter shown in FIG. 10b applied to image S is equivalent to a 7.times.7 low pass filter applied to Z. The 9.times.9 pixel low pass filter shown in FIG. 10c, applied to image M is equivalent to a 15.times.15 low pass filter applied to Z.

The moving average technique can be employed with the multi stage method described with reference to FIG. 9 as follows. Instead of processing the Z-S, S-M, and M-L images through SVD processes 40, each of the bandpass images can be processed through the block overlap SVD processing as described with reference to FIGS 7 and 8. The flow chart of this combined processing is shown in FIG. 11. Referring to FIG. 11, a group of filters and difference amplifiers 94 is employed to generate the bandpass and low pass images Z-S, S-M, M-L, and L from image Z as was described in FIG. 9. A block-overlap SVD process 96, 96', and 96", as shown in FIG. 7, is applied to bandpass signals Z-S, S-M, and M-L respectively to produce processed bandpass images Z-S', S-M', and M-L'. The processed bandpass images are summed (98) with low pass image L to produce processed image Z'.

To improve the response to diagonal edges having .+-.45.degree. orientation, the image blocks are sampled in a trapezoidal pattern as shown in FIG. 13. In this figure we show by example how a 4.times.4 block of pixels P.sub.n,m can be sampled in three different grid orientations, -45 (shown by dotted lines). 0 (shown by solid lines), and +45 degrees (shown by dashed lines). Note that the three differently oriented sampling patterns have a sub-block of 4 common pixels (P.sub.4,2 P.sub.5,2 P.sub.4,3 and P.sub.5,3) in the center. As shown FIG. 12, the SVD transform is performed on each of the three blocks of different orientations. The image detail signal H, generated as shown in FIG. 4, is processed by a 0.degree. block SVD transform 44, a +45.degree. block SVD transform 44', and a -45.degree. block SVD transform 44", to produce sets of singular vector matrices U and V.sup.T and singular values d.sub.i. The singular values d.sub.i are noise normalized (100, 102, 104) according to:

where i=1, 2, . . . n; and .mu..sub.i and .sigma..sub.1, are the mean and standard deviations, respectively, of the noise singular values d.sub.i generated for each block orientation (0.degree., 45.degree., and 45.degree.) as described previously with reference to FIG. 5. The block orientation most closely corresponding to the orientation of an edge is selected (106) on the basis of the values of the noise normalized singular value z.sub.i. Generally, if the orientation of the SVD blocks transform corresponds to the orientation of an edge in the image, its noise normalized singular values z.sub.i decrease faster as the index i increases, and the first few values z.sub.1 and z.sub.2 are higher than those of blocks not oriented with an edge. The selection is performed as follows. Starting from i=1, z.sub.i from each block is compared with the noise level say 3.5.sigma..sub.i for the block orientation. The lowest index i where the normalized singular value falls within the noise level is noted and denoted i.sub.n. If i.sub.n for all 3 block orientations (i.e. 0.degree., 45.degree., and -45.degree.) are different, the block orientation with the lowest i.sub.n is selected. If the lowest i.sub.n is the same for two orientations, then the orientation with the largest value of (z.sub.1 +z.sub.2) is selected. If i.sub.n is identical for all three orientations and i.sub.n is 1, then the region is most likely dominated by noise and the 0.degree. orientation is selected. If i.sub.n is the same for all three orientations and i.sub.n is equal to 2 then the orientation with the largest value of z.sub.1 is selected. If i.sub.n is the same for all three orientations and i.sub.n is greater than 2, then an orientation with the largest (z.sub.1 +z.sub.2) is selected. An orientation chosen in this way provides the best representation of the local image detail. A look up table 48 of factors f.sub.i for modifying the singular values is prepared as was described with reference to FIG. 5 for each block orientation. The appropriate factors f.sub.i are applied (46) to the singular values d.sub.i from the block having the selected orientation to produce modified singular values d.sub.i '. The modified singular values d.sub.i ' and singular vector matrices U and V.sup.T are inverse transformed (50) to produce a noise reduced image detail signal. The 4 pixels (P.sub.4,2 P.sub.5,2 P.sub.4,3 and P.sub.5,3) common to all three block orientations are extracted from each processed block to produce the noise reduced image detail signal H'. Finally, the noise reduced image detail signal H' is added back to the low pass image signal G to form the processed image signal Z' as shown in FIG. 4. Although the diagonal block SVD processing method was described with reference to 4.times.4 block of pixels for ease of description, the presently preferred block size is 20.times.20 pixels, having a region of 10.times.10 pixels common to all three blocks orientations.

The diagonal block SVD processing method is preferably implemented using the multistage block overlap technique described with reference to FIG. 11, and the center common block portion of 4 pixels is extracted from the SVD processed block of that orientation.

The method of the present invention can be applied to processing digital color images in a variety of ways. In one mode, the digital color image is separated into red, green, and blue color separation images, and each color separation image is processed using a block SVD transform method as described above. In the preferred mode of practicing the invention with a color image, the red, green, and blue color image is transformed into a luminance Y (e.g. Y=3/8R+4/8 G+1/8B) and two color difference components (R-Y and B-Y). Each of these images is processed using any one of the SVD block transform methods described above to produce processed luminance and color difference signals. The processed luminance and color difference signals are recombined to produce the processed color image In an alternative method, the red, green, blue color images are transformed into a luminance (Y), and two color difference components (R-Y and B-Y). Only the luminance image Y is processed using the block SVD transform method and the processed luminance image is recombined with the color difference components to produce a processed color image This approach reduces the processing time compared with the method discussed immediately above.

The image processing method according to the present invention is useful in graphic arts digital image processing and photographic digital image processing. The method has the advantages of producing a processed image having reduced noise and free from undesirable artifacts in areas of texture. The method has the further advantage that image detail is not degraded by the image processing. ##SPC1##