JPH10322702A - Picture decoding method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently remove a block distortion contained in a picture signal and to improve the quality of a picture in the ratio of signal to noise. SOLUTION: A code is decoded from encoding data 102 in a code decoding part 103 and a quantization conversion part 104 calculates the quantization representative value and the quantization clipping range of a conversion coefficient. The quantization representative value is converted in an inverse orthogonal conversion part 105 and the decoding picture signal is obtained. The picture signal is restored in a maximum posteriori probability estimation part 106 and the conversion coefficient is obtained in an orthogonal conversion part 108 through a conversion judgment part 107. The conversion coefficient is compared with the quantization clipping range in a coefficient comparison part 109. When it is settled in the quantization clipping range, the conversion coefficient is returned to the inverse orthogonal conversion part 105 as it is. When it is not settled in the quantization range, the coefficient settled in the quantization range is obtained in a random number generator 110 and it is returned to the inverse orthogonal conversion part 105.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a high-efficiency coding / decoding method using orthogonal transform based on small blocks.

[0002]

2. Description of the Related Art The main function of a conventional block distortion removing method is to apply a low-pass filter to an image signal and to blur boundaries between small blocks in which block distortion appears as a result.

Further, the projection onto a convex hull described below (POC
As for the repetition method using S), basically only a method using a low-pass filter has been proposed.

[0004] Here, the POCS theory will be briefly described. Image signal of Hilbert space element vector

[0005]

[Outside 1] Think. M convex hull sets in this space (each
The set corresponding to the requirement that the natural image signal should satisfy) is C _i
(I = 1, 2,..., M), and let P _i be the projection on it, any initial image signal

[0006]

[Outside 2] Vector column starting from

[0007]

[Outside 3]

[0008]

(Equation 1) Is

[0009]

(Equation 2) Converges to Here, I is a unit matrix, and λ _i is 0 <λ _i <
It is a real number that satisfies 2. Obviously image signal

[0010]

[Outside 4] Satisfies all the m types of requirements corresponding to C _i , and can be expected to have properties close to those of the original image.

In the repetition of the POCS, the image signal is gradually blurred only by repeatedly applying the low-pass filter,
There is a possibility that the image will be far from the original image. In order to avoid this, the quantized and coded orthogonal transform coefficients received by the decoder are used to monitor whether the restored image signal gradually departs from the original image signal.

That is, when the coefficient obtained by performing the orthogonal transformation on the image under restoration is not included in the corresponding quantization section, the coefficient is corrected (clipping) so as to be included. Conventionally, coefficients obtained by performing orthogonal transformation on an image being restored are corrected to the closer one of the boundaries of the section. By this operation, even if the restoration process is repeated, the processed image does not leave the original image without limit. One of these methods is Zakhor's method (A. Zakhor: "Iterativeprocedures for reduction of
blocking effects in transform image coding ", IEEE
Trans. CSVT., Vol.2, No.1, pp.91-95, Mar. 1992
).

Further, Kim et al. Has proposed a method of improving the signal-to-noise ratio of a restored image by narrowing the clipping section (DSKim et al .: "Projection").
ontothe narrow quantization constraint set for pos
tprocessing of scalar quantized images "SPIE Vol.2
727, pp. 1473-1483, 1996).

[0014]

When a low-pass filter is applied to an image signal as it is, block distortion is removed, and at the same time, a detailed portion of an image and an edge signal near a block boundary which should not be blurred are reduced. There was an undesired characteristic of being obscured.

In addition, there is a problem that the processing becomes complicated when the method of performing the texture analysis and locally switching the processing is performed.

Further, according to Zakhor's method of simply clipping the orthogonal transform coefficients in the quantization interval, the expected value of the square error between the original image and the restored image cannot be statistically minimized. Kim
Narrowing the clipping section as in these methods is not optimal processing.

An object of the present invention is to provide an image decoding method and apparatus which eliminates block distortion and further improves image quality in terms of a signal-to-noise ratio compared with a restored image.

[0018]

According to an image decoding method of the present invention, an image signal is divided into small blocks, and an orthogonal transform coefficient obtained by performing an orthogonal transform for each small block is defined as a quantization step width. To obtain a quantized representative value, decode the code assigned to the quantized representative value and record it to reproduce the quantized representative value, and inversely transform the reproduced quantized representative value to decode In an image decoding method for obtaining an image signal, the obtained image signal is restored with an image signal by a maximum a posteriori probability estimation method, and then the orthogonal image is orthogonally transformed to obtain an orthogonal transformation coefficient again. Then, the range of the orthogonal transform coefficient is specified from the code corresponding to the quantized representative value, and the coefficient is corrected so that the obtained orthogonal transform coefficient falls within the section, by an appropriate number of times according to the image.

According to another image decoding method of the present invention, an image signal is divided into small blocks, and orthogonal transformation coefficients obtained by performing orthogonal transformation for each small block are quantized by a predetermined quantization step width. A decoding step of obtaining a quantized representative value and decoding a code assigned and recorded to the quantized representative value;
A quantization transformation step of calculating a quantization representative value and a quantization clipping range from the decoding result of the code decoding step; an inverse orthogonal transformation of the quantization representative value to obtain a decoded image signal; and A maximum a posteriori estimation step of restoring the image signal by the maximum a posteriori estimation method for the image;
Orthogonal transformation of the restored image signal, an orthogonal transformation step of obtaining a transformation coefficient, comparing the transformation coefficient with the quantization clipping range, and performing the inverse orthogonal transformation step if the transformation coefficient is within the quantization clipping range. A coefficient comparison step of returning to the above, and in the coefficient comparison step, when the transformation coefficient does not fall within the quantization clipping range, obtain a transformation coefficient that falls within the quantization clipping range, and return to the inverse orthogonal transformation step. When it is determined that the above processing is repeated a predetermined number of times, or that the estimated maximum posterior probability of the original image signal has stopped decreasing or that it is increasing, a convergence determining step of outputting a restored image signal is provided.

The image decoding apparatus according to the present invention divides an image signal into small blocks, and quantizes the orthogonal transform coefficients obtained by performing orthogonal transform for each small block with a predetermined quantization step width. Code decoding means for obtaining a representative value and decoding a code assigned to the quantized representative value and recorded; and a quantization conversion means for calculating a quantized representative value and a quantization clipping range from the decoding result of the code decoding means. Inverse orthogonal transform of the quantized representative value to obtain a decoded image signal; an inverse posterior transform estimating unit for restoring an image signal to the decoded image by a maximal posterior definite estimation method; Orthogonal transformation of the restored image signal, orthogonal transformation means for obtaining a transformation coefficient, comparing the transformation coefficient with the quantization clipping range, and if the transformation coefficient is within the quantization clipping range, the transformation coefficient is Coefficient comparison means for returning to the orthogonal transformation step, and in the coefficient comparison means, when the transformation coefficient does not fall within the quantization clipping range, obtain a transformation coefficient that falls within the quantization clipping range, and the inverse orthogonal transformation means Returning means and a convergence determining means for outputting a restored image signal when it is determined that the above processing is repeated a predetermined number of times, or that the maximum posterior probability estimation value of the original image signal has stopped decreasing, or that it is increasing. Have.

According to the present invention, by repeatedly performing image restoration by estimating the maximum a posteriori probability and image restoration by monitoring the quantization range of the orthogonal transform coefficient, block distortion and the like included in the image signal can be efficiently removed. At the same time, the image quality is improved in terms of the signal-to-noise ratio.

In the repetitive processing, it is desirable to perform image signal processing in which edge portions that are not desirable to be removed are left as much as possible and block distortion is removed. Also, at this time, performing image local classification and performing adaptive processing for each region complicates the processing and increases the processing time. Therefore, it is desired that the entire image can be uniformly processed.

In addition, the clipping of the orthogonal transform coefficient is performed in view of the conversion distribution of the actual image signal, so that the optimum clipping is performed in order to further reduce the square error compared to the conventional method.

In the present invention, as the iterative image projection in the POCS, two methods are used, the maximum posterior probability estimation and the improved quantization interval restriction.

First, the estimation of the maximum posterior probability will be described. This method is a restoration method that projects the signal with noise superimposed in the direction with the highest probability.When applied to an image signal, unlike a low-pass filter, the parameters of the image model are set appropriately. By doing so, it is possible to leave an appropriate amount of edges.

The original image signal is

[0027]

[Outside 5] Restore the restored image signal

[0028]

[Outside 6] And

[0029]

[Outside 7] The maximum posterior probability of

[0030]

[Outside 8] Is the log likelihood function

[0031]

[Outside 9] Using

[0032]

(Equation 3) Becomes This is transformed by Bayes law,

[0033]

[Outside 10] Is irrelevant to maximization

[0034]

(Equation 4) Get. Where the image signal

[0035]

[Outside 11] Assuming that is an isotropic generalized Gauss-Markov random field,

[0036]

[Outside 12] Is

[0037]

(Equation 5) Becomes Here, α, b _{s, r} are constants, C is a set of pixel neighborhoods, s is any pixel of interest, and r is its neighboring pixels.

In the equation (7), p and q are parameters of the random field. When both are 1 or more, the objective function of the equation becomes convex. As p and q are smaller, the edge signal of the image is better preserved, and as p and q are larger, the edge is suppressed. If α is small, the pixel value is searched in a range that does not greatly depart from the start image. Since the present invention also uses the quantization interval restriction and monitors that the image does not leave the original image without any restriction, this may be arbitrarily large. Further, in order to remove block distortion, it is necessary to increase the penalty of the edge part to some extent. Therefore, both p and q are set to 2 in the present invention. This will moderate the edges,
Also, the slope of equation (7) is

[0039]

(Equation 6) , Which contributes to faster computation.

Next, an improved quantization section limiting method according to the present invention will be described. When observing the orthogonal transform coefficients and the true orthogonal transform coefficients of the image signal during the actual restoration, even coefficients that are not included in the quantization section are randomly distributed in the section almost without bias. Therefore, in the present invention, instead of the conventional simple clipping of the orthogonal transform coefficients, the transform coefficients outside the section are returned into the section by a random number generator. In this case, a sufficiently good result can be obtained even with a simple uniform distribution.

[0041]

Next, embodiments of the present invention will be described with reference to the drawings.

The image decoder of the present embodiment has a receiving terminal 101
, Decoding unit 103, quantization transformation unit 104, inverse orthogonal transformation unit 105, maximum posterior probability estimation unit 106, and convergence determination unit 10.
7, an orthogonal transformation unit 108, a coefficient comparison unit 109, a random number generator 110, and an output terminal 112.

After the code is decoded by the code decoder 103 from the coded data 102 received at the receiving terminal 101, the quantization converter 104 calculates the quantization representative value of the transform coefficient and the quantization clipping range. I do. The quantized representative value is transformed by the inverse orthogonal transform unit 105, and a decoded image signal on which noise such as block distortion is superimposed is obtained. The decoded image signal is restored by the maximum a posteriori probability estimating unit 106, passes through the convergence determining unit 107, and is again subjected to orthogonal transform by the orthogonal transform unit 108, thereby obtaining a transform coefficient. This conversion coefficient is compared by the coefficient comparison unit 109 with the quantization clipping range from the quantization conversion unit 104. If the transform coefficient falls within the quantization clipping range, the transform coefficient is directly returned to the inverse orthogonal transform unit 105. If not, the random number generator 110 obtains a coefficient that falls within the quantization range, and returns it to the inverse orthogonal transform unit 105. If the quantization clipping range is, for example, 8.5 to 11.5, (8.5 + 3r) is output using a uniform random number r of 0 to 1. This results in a uniform random number between 8.5 and 11.5.

While repeating this process, the convergence determination unit 107 determines that the number of iterations exceeds a predetermined number, that the value of the equation (7) has stopped decreasing, or that it is increasing. If so, the process ends, and the restored image signal 111 is finally output to the output terminal 112. Equation (7) is a measure for measuring the so-called naturalness of an image under the assumption that the image is an isotropic generalized Gauss-Markov random field, and the smaller this is, the more natural the image is.

[0045]

As described above, according to the present invention,
By repeatedly performing image restoration by estimating the maximum posterior probability and image restoration by monitoring the quantization range of the orthogonal transform coefficient, block distortion and the like included in the image signal are efficiently removed, and at the same time, the signal-to-noise ratio is reduced. Finally, the decoding side can obtain a high-quality image closer to the original image than a decoded image obtained by simple inverse orthogonal transform on the decoding side.

[Brief description of the drawings]

FIG. 1 is a block diagram of an image decoding device according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 101 Receiving terminal 102 Received encoded data 103 Decoding unit 104 Quantization transformation unit 105 Inverse orthogonal transformation unit 106 Maximum posterior probability estimation unit 107 Convergence judgment unit 108 Orthogonal transformation unit 109 Coefficient comparison unit 110 Random number generator 111 Final Restored image signal 112 output terminal

 ──────────────────────────────────────────────────の Continuing on the front page (72) Inventor Hiroshi Kodera Nippon Telegraph and Telephone Corporation, 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo

Claims (3)

[Claims] An image signal is divided into small blocks, and orthogonal transformation coefficients obtained by performing orthogonal transformation for each small block are quantized with a predetermined quantization step width to obtain a quantized representative value. An image decoding method for decoding a code assigned to a quantized representative value to reproduce a quantized representative value, and performing an inverse orthogonal transform on the reproduced quantized representative value to obtain a decoded image signal. For the image signal, perform restoration of the image signal by the maximum posterior probability estimation method, and then perform orthogonal transformation on the restored image signal to obtain an orthogonal transformation coefficient again, from the code corresponding to the quantized representative value. An image decoding method characterized in that a range of orthogonal transform coefficients is specified, and the coefficients are corrected so that the obtained orthogonal transform coefficients fall within the section, an appropriate number of times according to the image. 2. An image signal is divided into small blocks, and an orthogonal transform coefficient obtained by performing orthogonal transform for each small block is quantized with a predetermined quantization step width to obtain a quantized representative value. A code decoding step of decoding a code assigned and recorded to the quantization representative value; a quantization conversion step of calculating a quantization representative value and a quantization clipping range from a decoding result of the code decoding step; Inverse orthogonal transformation of the values to obtain a decoded image signal; a maximum a posteriori estimation step of restoring an image signal to the decoded image by a maximum a posteriori estimation method; and an orthogonal transformation of the restored image signal. And an orthogonal transformation step of obtaining transformation coefficients, comparing the transformation coefficients with the quantization clipping range,
A coefficient comparing step of returning the transform coefficients to the inverse orthogonal transform step if the transform coefficients fall within the quantization clipping range; and, in the coefficient comparing step, when the transform coefficients do not fall within the quantization clipping range, the quantization is performed. Obtaining a transform coefficient that falls within the clipping range, returning to the inverse orthogonal transform step, and repeating the above processing a predetermined number of times, or stopping the decrease of the maximum posterior probability estimate of the original image signal, or increasing the And a convergence determining step of outputting a restored image signal when the image decoding method is determined. 3. An image signal is divided into small blocks, and an orthogonal transformation coefficient obtained by performing orthogonal transformation for each small block is quantized with a predetermined quantization step width to obtain a quantization representative value. Code decoding means for decoding a code assigned and recorded to the quantization representative value; quantization conversion means for calculating a quantization representative value and a quantization clipping range from the decoding result of the code decoding means; Inverse orthogonal transformation means for inversely transforming a value to obtain a decoded image signal; maximum posterior definite estimation means for restoring an image signal to the decoded image by a maximum posterior definite estimation method; and orthogonal transformation of the restored image signal And orthogonal transform means for obtaining transform coefficients, comparing the transform coefficients with the quantization clipping range,
Coefficient comparing means for returning the transform coefficient to the inverse orthogonal transform step if the transform coefficient is within the quantization clipping range, wherein, when the transform coefficient is not within the quantization clipping range, the quantization Means for obtaining a transform coefficient that falls within the clipping range and returning to the inverse orthogonal transform means; repeating the above-described processing a predetermined number of times, stopping the decrease in the maximum posterior probability estimate of the original image signal, or increasing the An image decoding device having convergence determining means for outputting a restored image signal when the image decoding device determines that the image signal is in the first position.