Digital Image Noise Reduction Method and Transmission System

The present invention relates to a method for the restoring an image signal degraded by noise, and for a system for transmitting a restored image signal over a communication channel.

The goal of image restoration is to operate on an image signal that has been degraded by noise (such as film grain, scanner, or quantization noise) to remove the noise from the image signal and restore the image signal to its undergraded condition.

In one type of image restoration technique, statistical models of the image process and the noise process are constructed, and operations are performed on the noise degraded image signal to undo the effects of the noise process on the image signal. One such approach is to minimize a function of the error between the original image and the processed image. The operation on the image signal is called a minimum error filter. Minimum error filters can be derived from a priori and/or a posteriori knowledge about the statistics of the image and noise processes. A priori knowledge is that which is known before examining the degraded image signal, and a posteriori knowledge is that which can be determined by examining the noise degraded image signal itself.

A minimum mean-square error filter derived substantially from a priori knowledge is generally referred to as a Wiener filter. The minimum uniform error filter derived from a priori and a posteriori knowledge is generally referred to as a maximum a posteriori (MAP) filter. The MAP filter is a linear filter that maximizes the a posteriori probability for a linear imaging system.

The successful design of a MAP filter requires an estimate of the signal and noise probability density functions (pdf's) of the signal and noise. It has been shown that the MAP filter produces optimum results in the presence of uncertainty about the statistics of the image and noise processes. See H. J. Trussel, "Notes on Linear Image Restoration by Maximizng the A Posteriori Probability," IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP 26, No. 2, April 1978.

Unfortunately, the statistics of images vary considerably from image to image (i.e. the image statistics are not stationary) and although the linear MAP filter is more robust than other statistical filters in the presence of uncertainty about image statistics, the usefulness of the MAP filter is still limited due to the large variability in statistics from image to image.

It is an object of the present invention to provide a signal processing method for reducing noise in an image signal incorporating the advantages of a MAP filter that overcomes the shortcomings noted above due to the lack of stationary of the statistics of image signals.

In digital image transmission systems employing differential pulse code modulation DPCM (see DIGITAL IMAGE PROCESSING by William K. Pratt; John Wiley and Sons N.Y.; page 641) the difference between an actual pixel value, and its estimate, produced by a predictor, is quantized and coded for transmission over a communication channel. At a receiver, the decoded difference signal is reconstructed and combined with an estimate from a predictor identical to the one in the transmitter to reconstruct the original image. It is a further object of the present invention to incorporate the advantages of a MAP filter in a DPCM digital image transmission system to remove noise from the transmitted image signal.

The above objects are achieved according to one aspect of the present invention by providing a digital image transmission system having an encoder for converting a digital image signal to a differential pulse code modulated (DPCM) image signal. The encoder includes a differencing means for forming a DPCM image signal that is the difference between an incoming digital image signal and a prediction signal, a MAP filter that receives the DPCM image signal and produces a MAP filtered DPCM images signal, a quantizer that receives the MAP filtered DPCM image signal and produces a quantized DPCM image signal, and a feedback loop having a predictor for receiving the quantized DPCM image signal from the quantizer and generating and supplying the prediction signal to the differencing means. The MAP filter incorporates a MAP estimator for a signal having a Laplacian pdf, degraded by noise having a known pdf. The DPCM image signal is transmitted over a communications channel and decoded by a decoder having a summing means for forming an image signal representing the sum of the DPCM image signal and a prediction signal. A feedback loop including a predictor identical to the predictor in the transmitter receives the image signal from the summing means and generates and supplies the prediction signal to the summing means.

The above noted objects are achieved according to another aspect of the present invention by providing a method for reducing noise in an image signal degraded by noise having a known probability density function (pdf) including the steps of: sampling the image to produce a sampled image signal representing pixel amplitude values; digitizing the pixel amplitude values to produce a digital image signal; forming an image difference signal from the digital image signal by taking the difference between successive pixel amplitude values; processing the image difference signal in digital processing means including a digital MAP filter incorporating a MAP estimator for a signal having a Laplacian pdf degraded by noise having a known pdf, to produce a MAP filtered image difference signal; and recovering a restored digital image signal from the MAP filtered image difference signal by summing the amplitude difference values in the MAP filtered image difference signal.

The form of the MAP estimator employed in the invention is based on the empirical observation that the probability density functions of image difference signals for the types of images of interest to human observers are all very nearly Laplacian, and therefore highly stationary with only the variance changing from image to image. This observation also applies to the DPCM image signal produced in a DPCM image transmission system.

Since the pdf of the differential image signal is more nearly stationary than the image signal itself, a MAP estimator based on a differential image signal provides a more reliable estimate of the original signal statistics, even though the statistics of the original signal may vary widely.

When the MAP filter is inserted in the encoder of a differential pulse code modulation DPCM image transmission system noise correction and bandwidth compression are achieved in a single apparatus with minimum number of operations.

In preferred embodiments of the present invention, the noise to be removed from the signal is Gaussian, signal-dependent noise, such as film grain noise. A transformation is performed on the noisy image signal to render the noise independent of the image signal prior to filtering the differential image signal with the MAP filter, and the inverse transformation is performed on the restored image signal.

According to a further aspect of the invention, the image is segmented into regions of constant contrast by a generalized-likelihood-ratio-test on the variance of adjacent image pixel difference signals, and the variance of the respective segments is employed in the MAP filter.

FIG. 1 is a block diagram illustrating the noise reduction method according to the present invention;

FIG. 2 is a graph illustrating the uniform error cost function employed in the MAP estimation process;

FIG. 3 is a graph showing the adaptive, nonstationary coring effect of the MAP filter on an image difference signal;

FIG. 4 is a flow chart illustrating an image segment generator according to a preferred embodiment of the present invention;

FIG. 5 is a graph illustrating the functional form of predetermined constants employed in the generalized-likelihood-ratio-test used in the segment generator;

FIG. 6 is a flow chart illustrating a MAP filter according to a preferred embodiment of the present invention;

FIG. 7 is a graph illustrating a portion of an image signal useful in explaining the operation of the present invention; and

FIG. 8 is a block diagram illustrating a DPCM encoder and decoder including a MAP filter according to the present invention.

Before describing the signal processing method of the present invention, the derivation of the MAP estimate for a Laplacian signal degraded by signal-dependent Gaussian noise will first be described. The signal-dependent Gaussian noise model is a good model for noise such as film grain in a photographic image.

For a random signal (S) degraded by additive signal-dependent noise (n) the kth sample of the signal is given by

and the kth difference signal is given by ##EQU1##

The MAP estimator for .DELTA.S requires a local estimate of the variance of .DELTA.S. Since .DELTA.S is not available, its variance must be approximated by estimating the local variance of the observed signal .DELTA.X and reducing it by an estimate of the variance for the noise. If the noise is signal independent, reducing the variance involves straight forward subtraction of the noise variance from the observed signal variance. If however, the noise is signal-dependent, as is the case in most imaging processes, direct subtraction provides only an upper bound on the variance of .DELTA.S. A preferred approach is to perform a point-to-point transformation on the observed values of the image signal that makes the noise signal independent to first-order from the image signal, and then apply the inverse transformation after the noise has been removed from the signal.

The required transformation for additive signal-dependent Gaussian noise is ##EQU2## where C.sub.1 and C.sub.2 are arbitrary constants, and .sigma..sub.n (S') is the square root of the signal-dependent noise variance.

For film grain noise in a photographic image signal, .sigma..sub.n (S') is the Selwyn granularity as a function of photographic density, and is measured for a given film type by well known techniques. Although C.sub.1 is an arbitrary constant, the choice of C.sub.1 will determine the dynamic range of the transformed signal, and is therefore conveniently chosen so that the dynamic range of the signal matches the range of the A to D converters employed in the noise reduction apparatus as described below.

Designating transformed image signals with a tilde to distinguish them from the original signals, the kth sample of the transformed image signal is

and the corresponding transformed kth differential image signal is given by

For such a transformed differential image signal, the variance of .DELTA.S can be estimated to first-order by simply subtracting the transformed noise variance from the local variance estimates of the observed transformed signal.

In the following description, the tildes are omitted from the notation for simplicity, with the understanding that the inverse transformation will be performed on the processed signal to recover the original noise reduced signal. The MAP estimator for .DELTA.S is the most probable .DELTA.S given the observed noise degraded signal .DELTA.X. The MAP estimator corresponds to the minimum cost estimate for a cost function that is equal to 1 for incorrect estimates, and 0 otherwise. A plot of such a cost function is shown in FIG. 2. Symbolically, the MAP estimate .DELTA.S can be expressed in terms of probability density functions (pdf's) of .DELTA.S and .DELTA.X as: ##EQU3##

As noted above, it has been observed that the pdf of a differential image signal is modeled well by a Laplacian function, and assuming a Gaussian pdf for the now signal-independent noise, the distributions in equation (6) can be written as: ##EQU4## where K.sub.0 and K.sub.1 are normalization constants. The results are simplified if the natural logarithm of the pdf's are used rather than the pdf's themselves, and since the natural logarithm is a monotonic function, the substitution does not change the MAP estimate, which becomes: ##EQU5##

Since the last term in equation (9) is not a function of .DELTA.S, it can be dropped as can the normalization constants. Substituting equations (7) and (8) into equation (9), the MAP estimate becomes: ##EQU6## This equation is maximized for the two cases .DELTA.S.ltoreq.0 and .DELTA.S.gtoreq.0 by differentiating with respect to .DELTA.S and setting the result equal to zero as follows.

FIG. 3 is a plot of .DELTA.S vs. .DELTA.X generated by the MAP estimator defined by equation (17). As can be seen from FIG. 3, the MAP estimator has the effect of a coring operator on the image difference signal .DELTA.X. In the coring region of the function, shown by cross hatching, the estimated value of the image difference signal .DELTA.S is set to zero. In the regions outside the coring region, the estimated value of the image difference signal is biased toward zero by an amount directly related to the noise variance and that is inversely related to the square root of the estimated differential signal variance. The larger the noise variance or the smaller the image difference signal variance, the larger the bias. In the cored region, where the observed .DELTA.X values are smaller than ##EQU15## the signal .DELTA.S is set to zero, resulting in a significant smoothing in uniform low contrast areas of an image. The coring threshold is adaptive to the amplitude of the differential signal and to changes in the estimated variance of the differential signal and noise.

To implement the MAP filter, it is necessary to estimate .sigma..sub.n.spsb.2 and .sigma..sub..DELTA.S.spsb.2 which are defined as

where C.sub.1 is the constant employed in the transformation equation (3) above.

The unknown parameter .sigma..sub..DELTA.X.spsb.2 is estimated from local statistics of the image difference signal by a maximum-likelihood estimator (MLE) for the standard deviation of a Laplacian pdf given by ##EQU16## over a sliding window of fixed length (e.g. 10 samples).

Alternatively, the image difference signal may be segmented into variable length regions having stationary statistics. This is equivalent to segmenting the image signal into regions of constant contrast. The segmenting method, described below, is based on a generalized-likelihood-ratio-test (GLRT) for Laplacian data. The test statistically examines the hypothesis H.sub..phi. that the standard deviation of a given stationary segment is equivalent to the standard deviation of a second segment that is adjacent to the given segment and composed of a single image difference signal sample value. If the hypothesis is accepted, the new image difference signal sample value is added to the current segment, the standard deviation is updated, and the hypothesis is checked against the next sample value. If the hypothesis is rejected, the current segment is closed and the new sample value is assumed to be the first sample in a new segment. This method of segmenting images into statistically stationary segments is th subject of copending Patent Application No. 802,113 now U.S. Pat. No. 4,764,971.

The of processing an image signal with the MAP filter to reduce noise according to the present invention will now be described with reference to FIG. 1. The method shown in FIG. 1 can be performed in an appropriately programmed general purpose digital computer, or in hardware or firmware specifically designed for implementing the operations described below.

The image signal is sampled 20 by a sampling means such as a solid state image sensing array to produce a sampled image signal having pixel amplitude values representing luminesce, density or brightness. The sampled image signal X having signal dependent noise n(S) is transformed 22 by a transformation filter to render the noise independent of the signal. The transformation t(X) is described by equation (3) above. The transformed signal X is digitized 23 by a digitizer to produce a digital image signal. The digital image signal is segmented 24 in a segment generator that forms an image difference signal for each line of the scanned image, and segments the image difference signal into segments having stationary statistics, using a generalized-likelihood-ratio-test (GLRT) as described below. The segment generator produces signals S.sub.k representing the location of the start of the segments, the variance .sigma..DELTA.X.sub.k of the image difference signal in the segment, and the mean value .mu..sub.k of the image signals in the segment. The mean value of the image signal is employed in reconstructing the image signal as described below. The image difference signal is MAP filtered 30 in a MAP filter that receives the delayed image signal, the statistical information S.sub.k, .mu..sub.k, and .sigma..DELTA.X.sub.k and the variance of the noise .sigma..sub.n, supplied externally as described below, generates an image difference signal for each line of the image, and performs the MAP restoration on the noisy image difference and recovers a restored digital image signal from the MAP filtered image difference signal. The restored digital image signal is inverse transformed 32 in an inverse transformation filter to recover the noise restored signal S. For X=t(x), the inverse transform can be seen from equation (3) to be t.sup.-1 (X)=x. The step of forming an image difference signal is performed in both the generation step and the MAP filter step in the particular embodiment described. Alternatively, the image difference signal may be formed in a separate step and supplied to the segment generation step and the MAP filter step.

Detailed operation of the segment generator 24 will now be described with reference to the flow chart of FIG. 4. A person skilled in the art of computer programming will readily be able to produce a computer program for a general purpose digital computer to implement the segment generator described in FIG. 4. The following description assumes that the image signal is in the form of a raster scan. At the start of each scan line, the segment generator 24 initializes (100) internal counters: n indicating the number of pixels in a segment; j indicating the pixel number in the line; and k indicating the number of segments. The segment generator sets a pointer S.sub.k that points to the starting pixel for the kth segment, and initializes a sum .SIGMA. used to compute the mean value of the image signal X in a segment. The segment generator increments n and j by 1 (102) and then forms (104) a difference signal .DELTA.X.sub.n between the jth and j-1st samples of the image signal. A maximum likelihood estimate .sigma. of the variance of the difference signal .DELTA.X.sub.n as given by equation (20) is then computed (104).

Next, the segment generator checks (106) to see if this is the first pixel in a segment (n=1). If so, .DELTA.X.sub.1 is set (108) equal to .DELTA.X.sub.n, .sigma..sub.0 (an estimate of the variance of the segment as generated thus far) is set (110) to equal the square root of two times the absolute value of .DELTA.X.sub.1 and in is incremented by 1 (102). The next difference signal .DELTA.X.sub.n is formed (104), and the maximum likelihood estimate .sigma. of the variance of the difference signal is formed (104). This time, n does not equal 1, and the sum .SIGMA. is incremented by adding X.sub.j-1 (112). Next the null hypothesis (H.sub..phi.) that the sigma value of the present segment is equivalent to the sigma value of the present segment plus the next image difference signal sample is checked (114). If the null hypothesis is true, and it is not the end of the scan line (116), .sigma..sub.0 is updated (118) by factoring in .sigma., and another .DELTA. X is formed (104 and added to the segment for testing. If the null hypothesis is false, the segment is ended, k is incremented (120) by 1, and the values S.sub.k, .sigma..DELTA.X.sub.k-1 and .mu..sub.k-1 are stored (122) in a register. A check is made to see if this is the end of the line, if so, k is incremented by one (125), if not, .SIGMA. is updated (124), .DELTA.X.sub.1 is set to .DELTA.X.sub.n (108), .sigma..sub.0 is set to the square root of two times the absolute value of .DELTA.X.sub.1 (110), n is set to 1 (126), and the next segment building process is started. When the last pixel of a scan line is reached (end of line equal yes) k and n are incremented, sigma is updated by adding the last X value (128), S.sub.k is set to 0, .sigma..DELTA.X.sub.k-1 is updated by factoring in the most recent .sigma., .mu..sub.k-1 is set to .SIGMA./n, and these values are stored in the register (130,132) for the last segment of the line. This segmentation process has single pixel resolution because a new segment can start at any pixel and therefore the maximum number of segments possible is equal to the number of pixels in a line, and the storage register is sized accordingly.

The check on the truth of the null hypothesis H.sub.100 involves testing whether

where a and b are constants derived from estimation theory that depend on the value of n, the number of pixels in the segment. The equations for a and b for a Laplacian distribution are ##EQU17## respectively, where a is an empirically selected level of significance for the test. An equivalent statement of the decision rule is to reject H.sub..phi. ##EQU18## for a selected level of significance .alpha., the values of a and b vs. n are computed off-line using equations (23) and (24) and stored in a memory. FIG. 5 shows a plot of the dependence of a and b on n for a level of significance .alpha. of 0.25.

Turning now to FIG. 6, the MAP filter will be described in more detail. A person skilled in the art of computer programming will readily be able to produce a program for a general purpose digital computer to implement the MAP filter described in FIG. 6. Bearing in mind that the MAP filter has available to it the mean .mu..sub.k of the image signals in a segment and the variance .sigma..DELTA.X.sub.k of the image difference signals from the stationary segments in this line of image signals provided from the segment generation step 24.

The MAP filter initializes counters j and k (134), representing the pixel number and the segment number, respectively. Then the variance of the first segment .sigma..DELTA.X.sub.1 is checked (136) to see if it is greater than the square root of two times the variance of the noise .sigma..sub.n. If not, the first processed image value X is set equal to the mean value .mu..sub.k of the first segment (138), and the processed value X is output (140) as the first processed image value. If the variance of the first segment is greater than the square root of two times the the variance of noise, the first processed value X is set equal to the first signal value X.sub.0. The estimated variance .sigma..sub..DELTA.S of the first segment is then adjusted by the root square of the noise variance, and the next difference signal is calculated (142). The absolute value of the next difference signal .DELTA.X is tested (144) to determine whether it is in the adaptive coring region ##EQU19## If so, the value of X is replaced by the mean value .mu..sub.k for the segment (138). If not the first value X.sub.0 is output (140). Next, j is incremented (146) by one and a check (148) is made to determine if a new segment has been reached (j=S.sub.k+1 ?). If so, k is also incremented by one (150).

Next, the value of the variance of the signal .sigma..DELTA.X.sub.k is checked (152) to see if it is greater than the square root of two times .sigma..sub.n. If not, the value of the signal X.sub.j is replaced (154) by the mean value .mu..sub.k of the signal in the segment. Then a noise reduced image difference signal .DELTA.S is formed (156) by subtracting the previously determined signal value X from the jth sample X.sub.j. If on the other hand, the variance of the segment .sigma..DELTA.X.sub.k is greater (152) than the square root of two times .sigma..sub.n, a new estimate of the variance is computed by reducing the variance by the root square of the noise variance, and the next difference signal .DELTA.X is formed (158). The absolute value of the image difference signal is checked (160) to determine whether it is within the adaptive coring region ##EQU20## If so, the value of X is replaced If so, the value of X.sub.j is replaced by the mean value .mu..sub.k of the segment (154), and a noise reduced difference signal .DELTA.S is formed (156) by subtracting the last noise reduced image signal X from X.sub.j.

If the magnitude of the image difference signal lies outside of the coring region, the polarity of the image difference signal is tested (162) ##EQU21## and depending on whether it is positive or negative, a noise reduced image difference signal .DELTA.S is formed by respectively reducing (164) or increasing (166) the image difference signal .DELTA.X by ##EQU22## Finally, the new restored image signal X is recovered by adding (168) the noise reduced difference signal .DELTA.S to the previous value of X. The step of recovering the restored image signal is shown as part of the MAP filter, but may be performed as a separate step. The new value of the noise reduced image signal is output (170), and a check is made (172) to determine whether the end of the line has been reached. If not, the value of j is incremented by one (146), and the loop is repeated until the end of the scan line.

In the implementation described above, it will be noted that whenever .sigma..DELTA.X.sub.k was less than or equal to the square root of two times .sigma..sub.n, or when the absolute value of .DELTA.X was less than or equal to two times the square root of .sigma..sub.n.sup.2 /.sigma..DELTA.S the value of the signal was changed to the mean value of the signal in the segment. This has the effect of forcing the restored signal to have the average value of the signal in a area where there is some signal variance (.sigma..DELTA.X>0) but where the variance is not large enough to move the signal out of the coring region. The result is some additional smoothing of the noise restored image signal.

The signal processing method for reducing noise in an image signal will now be described with reference to a numerical example shown in the graph of FIG. 7. FIG. 7 shows the first seven samples in a line of image information. The sample values X.sub.j are shown above the graph. The level of significance .alpha. was chosen to be 0.25, i.e. 75% confidence that the rejection of the null hypothesis H.sub.100 is correct. A table of values for a and b for segments of given length n at a confidence level of 75% are shown in Table 1 below. These values were computed from equations (22) and (23) and stored in a table look-up memory.

The signal shown in FIG. 7 is segmented as shown by the values in Table 2 where .DELTA..sub.ij =.vertline.X.sub.i -X.sub.j .vertline..

The statistical values supplied to the MAP filter from the segmentation step are shown in Table 3.

Next, using the statistical values generated in the segmentation step, the values produced by the MAP filter are as shown in Table 4.

In the example given, the root mean-square error between the original signal S and the noise degraded signal X was ##EQU23## After restoration by the MAP filter, the RMSE error was ##EQU24## which represents approximately a 20% decrease in the root mean square noise in the signal.

In a DPCM digital image transmission system, the DPCM image signal that is transmitted over a communication channel is produced by forming the difference between an incoming image signal and a prediction signal produced in a feedback loop by a predictor, based on the previous value of the DPCM image signal. Like the image difference signal discussed above, the DPCM image signal is an image difference signal that also exhibits a Laplacian statistical distribution.

Since the MAP filter described above operates on a differential image signal having Laplacian statics, it can be implemented in the transmitter of a differential pulse code modulation (DPCM) image transmission system to remove noise in the image signal. FIG. 8 is a block diagram of a DPCM digital image transmission system including a MAP filter according to the present invention. The DPCM digital image transmission system includes an encoder generally designated 200, and a decoder generally designated 202.

The encoder 200 includes a sampler 204 for sampling an image signal to produce a digital image signal and a transformation filter 206 to make the noise in the signal independent of the signal. The transformation t(x) performed by transformation filter 206 is described in equation (3) above. The transformed digital image signal is supplied to a differencing node 208 where a prediction signal is subtracted from the digital image signal to produce a DPCM image signal. The DPCM image signal is supplied to a MAP filter 210, where it is filtered to remove noise. The noise filtered DPCM image signal is supplied to a Quantizer 212, where it is quantized. The quantized DPCM image signal is supplied to a channel coder 214, where it is coded for transmission over a communication channel 215. The quantized DPCM image signal is also supplied to a feedback loop 216 containing a predictor 218 that generates the prediction signal that is subtracted from the image signal to produce the DPCM image signal.

The predictor 218 supplies a prediction of zero at the start of each line in response to a line start signal 219 so that the first value in a line that is supplied to the MAP filter 210 is the actual image signal value rather than the DPCM image signal. The decoder 202 for receiving and decoding the DPCM image signal on the channel includes a channel decoder 220, a summing node 222 for summing the DPCM image signal and a prediction signal, a feedback loop 224 including a predictor 226 for forming the prediction signal from the sum of the DPCM image signal and the prediction signal, and an inverse transformer 228 to perform the inverse transform of the transformation performed by transformer 206 in the encoder on the recovered image signal.

As described above with respect to the signal processing method for reducing noise in an image signal, the image difference signal may be segmented into variable length regions having stationary statistics, and the statistics of the segments supplied to the MAP filter. Similarly, the the DPCM image signal may be segmented into regions of stationary statistics, and the statistics employed by the MAP filter 210 in the encoder 200. A segment generator 230 for generating the segments and their statics and supplying the statistics to the MAP filter 210 is shown by dashed lines in FIG. 8. To allow the segment generator 230 time to generate the statistical information before it is applied to the DPCM image signal, a one line delay 232 is provided at the input to differencing node 208. Like the MAP filter 210, the preferred implementation of the segment generator 230 is by a programmed microprocessor.

Sampler 204, differencing node 208, quantizer 212 channel coder 214, and feedback loop 216 including predictor 218, channel decoder 220, summing node 222, and feedback loop 224 including predictor 226 are well known in the prior art, requiring no further description. Transformer 206, MAP filter 210, and inverse transformer 228 are preferably implemented by a programming a microprocessor in a microprocessor based DPCM encoder/decoder board.

A Fortran program for implementing a DPCM encoder and decoder having a transformer to convert an input image with signal dependent noise to signal independent noise, an inverse transformer, a MAP filter, and an image segmentor is included in Appendix A.

The signal processing method for reducing noise in an image signal is useful for removing the noise from images such as photographic images and X-ray images. The method has the advantage that since the MAP restoration operates on the image difference signal which has more nearly stationary statistical properties than the image signal itself, the statistical assumptions that are required in the image restoration process are more nearly valid, and hence result in more accurate image restoration. The method has the further advantage that since the differential image data has a smaller dynamic range than the direct image data, fewer bits are required to express the image data, and the MAP filter operating on image difference data can be implemented in less complex hardware.

Furthermore, for a Laplacian image difference signal pdf and a Gaussian noise pdf, the MAP filter for the image difference signal behaves as an adaptive coring operator on the image difference signal with the only empirically derived parameter being the confidence level .alpha.. Because the coring operator is self adaptive and based on statistical criteria, it is more robust than prior art empirically derived coring operators, thereby making the method applicable to a wider class of images.

Since the MAP filter is linear in the uncored region and a null operator in the cored region, and since a DPCM image signal is an image difference signal with Laplacian statistics, the MAP filter is easily implemented in the transmitter of a DPCM data transmission system for transmitting and restoring images degraded by Gaussian noise. ##SPC1##