JP2015032924A - Conversion quantization method, conversion quantization device and conversion quantization program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect AZB and to adjust conversion units in different sizes by weighting.SOLUTION: From a conversion coefficient threshold table indicating a relation of a conversion quantization size, a QP value, a bit depth and a conversion coefficient threshold, a conversion coefficient threshold is determined which corresponds to a conversion quantization size N, an inputted QP value and a bit depth of a block in an N×N size (N=2and (n) is an integer equal to or greater than 3) subjected to image coding. A result of multiplying the conversion coefficient threshold by an adjustment coefficient determined for the conversion quantization size is defined as a differential absolute value sum upper limit value. A differential absolute value sum is compared with the differential absolute value sum upper limit value. In the case where it is determined that the differential absolute value sum is smaller than the differential absolute value sum upper limit value, the quantization conversion coefficient is made into zero. In the case where it is determined that the differential absolute value sum is greater than or equal to the differential absolute value sum upper limit value, the block in the N×N size is divided into four blocks in an (N/2)×(N/2) size and the determination is executed again.

Description

  The present invention relates to a transform quantization method, transform quantization apparatus, and transform quantization program for HEVC (High Efficiency Video Coding) using residual feature quantities.

  HEVC, which is attracting attention as a next-generation video compression standard, has excellent compression efficiency due to techniques such as optimization of the block size. In comparison with H.264 / AVC, it is announced that the bit rate is reduced by about 40%. However, while compression efficiency is excellent, compression is said to require several times the amount of calculation (see, for example, Non-Patent Document 1). This large amount of calculation results mainly from “tree structure prediction” and “multiple prediction modes”, which are features of HEVC.

  In order to solve the problem related to the large amount of calculation, an all-zero block (AZB) algorithm has been used in a conventional encoding technique. The AZB algorithm is a method in which an all-zero block, that is, a block in which all coefficients after quantization are all zero can be determined in advance without calculating transform quantization. Since it is not necessary to perform transform quantization on a block determined to be an all-zero block, the calculation amount is reduced as a result.

  Many of the proposals so far have been related to H.264, which is a conventional video compression standard. H.263 8 × 8 conversion and H.264. This is done for AZB detection in H.264 4 × 4 integer transform. In Non-Patent Document 2, the determination of all-zero transform coefficient blocks is not analyzed in detail, so that sufficient conditions are not derived and erroneous detection occurs. In Non-Patent Document 3, theoretical conditions are derived in order to improve the conventional algorithm. Non-Patent Document 4 proposes a low error detection determination method based on an analysis of necessary and sufficient conditions for early detection of all-zero coefficient blocks. In Non-Patent Document 5 and Non-Patent Document 6, H. A model applied to 4 × 4 integer conversion in H.264 / AVC is proposed. Recently, in Non-Patent Document 7, H.C. A more effective sufficient condition has been derived for the early determination of all-zero 4 × 4 blocks in H.264 / AVC.

Yoon Mi Hong, Kyo-Hyuk Lee Il-Koo Kim, Madhu Krishnan Wei Dai, and Pankaj Topiwala, '' Fast integer transforms for the HM, and complexity analysis, '' Joint Collaborative Team on Video Coding, JCTVC-D365, Daegu, 2011. Zhou Xuan, Yu Zhenghua, and Yu Songyu, '' Method for detecting all-zero dct coefficients ahead of discrete cosine transformation and quantisation, '' Electronics Letters, vol. 34, no. 19, pp. 1839 --1840, sep 1998 . LA Sousa, '' General method for reducing redundant computations in video coding, '' Electronics Letters, vol. 36, no. 4, pp. 306--307, 2000. S. Jun and S. Yu, `` Efficient method for early detection of all-zero DCT coefficients, '' Electronics letters, vol. 37, no. 3, pp. 160--161, 2001. Y. Wan, Y. Zhou, and H. Yang, `` Early detection method of all-zero integer transform coefficients, '' Consumer Electronics, IEEE Transactions on, vol. 50, no. 3, pp. 923--928, 2004. YH Moon, GY Kim, and JH Kim, `` An improved early detection algorithm for all-zero blocks in H. 264 video encoding, '' Circuits and Systems for Video Technology, IEEE Transactions on, vol. 15, no. 8, pp. 1053--1057, 2005. H. Wang, S. Kwong, and CW Kok, Efficient prediction algorithm of integer DCT coefficients for H.264 / AVC optimization, '' Circuits and Systems for Video Technology, IEEE Transactions on, vol. 16, no.4, pp. 547--552, 2006.

  As described above, many AZB algorithms based on the conventional video coding standard have been proposed, but there is a problem that a sufficient effect cannot be obtained only by directly applying them to HEVC. The reasons are mainly the following two points.

  The first point is that in the quadtree structure adopted in HEVC, the AZB algorithm must be applied for both the block size and depth. The depth is a concept representing the number of divisions of a block divided by a quadtree structure, and it is necessary to apply an algorithm corresponding to both various sizes and the division depth.

  The second point is that by applying the AZB algorithm, as a result, a large-sized transform unit (TU: Transform Unit) tends to be selected and terminated. It is necessary to make adjustments for various sizes of TUs.

  The present invention has been made in view of such circumstances, and makes it possible to detect AZB for various block sizes, and to enable conversion units having different sizes to be adjusted by weighting. It is an object to provide a quantization method, a transform quantization apparatus, and a transform quantization program.

The present invention relates to a transform quantization method for calculating a quantized transform coefficient from an inter prediction residual in image coding and a sum of absolute differences thereof, and the relationship between transform quantization size, QP value, bit depth, and transform coefficient threshold value Corresponding to the transform quantization size N, the input QP value, and the bit depth of an N × N size block (N = 2 ⁿ , where n is an integer of 3 or more) that is the object of image coding. A transform coefficient threshold specifying step for obtaining a transform coefficient threshold; and a difference absolute value sum upper limit calculating step in which a result obtained by multiplying the transform coefficient threshold by an adjustment coefficient determined for the transform quantization size is a difference absolute value sum upper limit value; The difference absolute value sum and the difference absolute value sum upper limit value are compared with each other, and the AZB determination step determines that the difference absolute value sum is smaller than the difference absolute value sum upper limit value. In the case where it is determined that the difference absolute value sum is greater than or equal to the difference absolute value sum upper limit value by the quantization transform coefficient determination step in which the quantization transform coefficient is zero and the AZB determination step Has a dividing step of dividing the N × N size block into four (N / 2) × (N / 2) size blocks and executing the AZB determination step again.

  According to the present invention, when the inter prediction residual and the transform coefficient follow a Gaussian distribution, and the absolute values of the quantized transform coefficients of the block to be image-coded are all smaller than 1 in the transform coefficient threshold table, The value derived in the above is defined as a conversion coefficient threshold value.

The present invention is a transform quantization apparatus that calculates a quantized transform coefficient from an inter prediction residual in image coding and a sum of absolute differences thereof, and the relationship between transform quantization size, QP value, bit depth, and transform coefficient threshold value Corresponding to the transform quantization size N, the input QP value, and the bit depth of an N × N size block (N = 2 ⁿ , where n is an integer of 3 or more) that is the object of image coding. A transform coefficient threshold specifying means for obtaining a transform coefficient threshold; and a difference absolute value sum upper limit calculating means for setting a difference absolute value sum upper limit as a result of multiplying the transform coefficient threshold by an adjustment coefficient determined for the transform quantization size; The difference absolute value sum and the difference absolute value sum upper limit value are compared, and the AZB determination means determines that the difference absolute value sum is smaller than the difference absolute value sum upper limit value. Is When the quantized transform coefficient determining means for setting the quantized transform coefficient to zero and the AZB determining means determine that the sum of absolute differences is greater than or equal to the upper limit of absolute differences, the N × A dividing unit that divides the N-sized block into four (N / 2) × (N / 2) -sized blocks and performs the determination by the AZB determining unit again.

  The present invention is a transform quantization program for causing a computer to execute the transform quantization method.

  According to the present invention, it is possible to perform optimum transform quantization by making it possible to detect AZB for various block sizes, and enabling adjustment by weighting for transform units of different sizes. The effect that it becomes possible is obtained.

It is a block diagram which shows the structure of the encoding process part of HEVC by one Embodiment of this invention. It is a figure which shows the AZB determination process by a prior art. It is a figure which shows the AZB determination process by this embodiment. It is a figure which shows the process of conversion quantization calculation. It is a figure showing the derivation process of an AZB threshold value. It is a figure which shows the energy after calculating the same TU with a different conversion size. It is a figure which shows an AZB detection rate.

  Hereinafter, a transform quantization apparatus according to an embodiment of the present invention will be described with reference to the drawings. The proposal according to the present invention is composed of two techniques, one is AZB detection (variable block size AZB detection) for various block sizes, and the other is weight adjustment (adjustment by weight). In order to detect AZB of various block sizes, the residuals and transformed coefficients are modeled assuming a Gaussian distribution. The variance is calculated assuming a Markov process. In order to become AZB, all the coefficients need to be smaller than the DC component of the covariance, so AZB is determined using this condition. Further, by adjusting the weight, the weight for the larger size conversion is reduced. By doing this, a smaller conversion can be considered.

FIG. 1 is a block diagram showing a configuration of an encoding processing unit of HEVC according to the embodiment. The encoding processing unit includes an inter prediction unit 1, an AZB determination unit 2, an integer transform quantization unit 3, and an entropy coding unit 4. After the inter prediction is processed, the residual r _ij and its SAD (sum of absolute differences) are sent to the AZB determination unit 2 in order to determine whether or not conversion / quantization calculation is required. The SAD value is sent from the inter prediction unit 1 to eliminate the need for recalculation by the AZB determination unit 2. If the block is determined as AZB in the AZB determination unit 2 according to the AZB determination method, the transform / quantization matrix q _ij can be directly set to 0 and transmitted to the entropy encoding unit 4. The value of QP (quantization parameter) is input to the inter prediction unit 1, the AZB determination unit 2, the integer transform quantization unit 3, and the entropy coding unit 4.

  Next, the AZB determination process will be described. First, the AZB determination process according to the conventional method will be described with reference to FIG. FIG. 2 is a diagram illustrating an AZB determination process according to the prior art. As shown in FIG. 2, in the conventional method, threshold calculation is performed for a specific block size (4 × 4 blocks in the figure) (step S1), and it is determined whether or not the value of SAD is equal to or less than the threshold (step S2). By doing so, it is determined whether or not it is an all-zero block.

  Next, the method according to the present embodiment will be described with reference to FIG. FIG. 3 is a diagram illustrating AZB determination processing according to the present embodiment. This method can perform efficient AZB determination even in HEVC encoding in which a plurality of block sizes are candidates by continuously performing threshold determination from a large size to a small size.

The conversion quantization size (N) and the QP value (QP) are input, and the threshold value table that can output the bit depth (Bit depth) is referred to (steps S11, S12, and S13), and the input conversion quantization is performed. A threshold (bit depth) for the size (N) and the QP value (QP) is obtained. Examples of threshold tables are shown in Tables 1 and 2. It is assumed that the threshold values shown in Tables 1 and 2 are calculated in advance.

  In the inter prediction process, the SAD is calculated by motion estimation for one coding unit (CU). The SAD value in the currently input residual block size is compared with the corresponding threshold value (step S14). If the SAD is smaller than the threshold value, the current residual block (for example, 2N × 2N) is recognized as an all zero block (2N × 2N). If the condition is not satisfied, the residual block is further divided into smaller sizes and repeatedly determined (steps S14 and S15).

In the example shown in FIG. 3, the processing order is performed from a large size to a small size, so that a block of a larger size is checked in advance. As a result, the method according to the present embodiment has an effect of determining AZB with a larger size. That means that even when the AZB occurs with a smaller TU size, it is ignored even if it is the optimal solution. In order to overcome this deficiency, the weighting threshold values (W _2N TH (QP) _{2N × 2N} , W _N TH (QP) _{N × N} , W _{N / 2} TH (QP) _{N / 2 × N / 2} ) are used. Apply. W _2N , W _N and W _{N / 2} represent the block size weights of the residuals corresponding to (2N × 2N, N × N, (N / 2) × (N / 2)), respectively. TH (QP) _{2N × 2N} , TH (QP) _{N × N} and TH (QP) _{N / 2 × N / 2} are respectively the corresponding residual block size thresholds (2N × 2N, N × N, (N / 2) × (N / 2)).

From the above, the difference between the method according to the present embodiment shown in FIG. 1 and the conventional method can be summarized as follows.
(1) A threshold table corresponding to each TU size is calculated and used in advance in order to perform the determination process at high speed.
(2) Weighting adjustment is added to eliminate the effect of determining AZB at a larger size.

  Next, detailed derivation of the AZB threshold will be described. First, derivation of AZB threshold values for different block sizes will be described. FIG. 4 is a diagram illustrating a process of transform quantization calculation. As shown in FIG. 4, the derivation of the threshold follows a reverse of the transform quantization calculation process. That is, after transforming from the residual, quantization is performed to obtain a coefficient.

FIG. 5 is a diagram illustrating the derivation process of the AZB threshold value. If the absolute values of all quantized transform coefficients (quantized transform coefficients q _ij ) are less than 1, the block is AZB. In this case, the transform coefficient before the quantization (transform coefficient d _ij ) is limited to a range in which the quantization parameter (QP) is distributed. Further, the relationship between the value of the input residual signal and the transformation coefficient is modeled as a Gaussian distribution. By calculating the difference between the Gaussian distribution of the input residual signal and the Gaussian distribution of the transform coefficients, the relationship between the sum of absolute differences (SAD) and the AZB threshold is constructed.

Based on the scaling and quantization definitions in HM6.3 (HEVC test model version 6.3), the quantized transform coefficients q _ij for a particular N × N block are ((i, j = 0.. , N−1), and the transform coefficient d _ij (i, j = 0... N−1) before quantization is calculated as follows.
here
And BitDepth is the bit depth of each pixel. If | q _ij | <1, this block is an all-zero block.

That is

therefore,
Is satisfied and the following equation is established.

Hereinafter, TH _TR [QP] _ij will be referred to as a conversion coefficient threshold value. Since the residual sample r _ij (i, j = 0... N−1) and the pre-quantization transformation coefficient d _ij which are inter prediction residuals are modeled as a Gaussian distribution, the distribution of d _ij is (− nσ, nσ). n is a real number greater than or equal to zero. When n = 3, the probability that _dij is in this range is approximately 99%. Then, when a relationship (approximate expression) established between the distributions of the two distributions when the distribution σ _T (u, v) of the transform coefficient and the distribution σ _r of the inter prediction residual is a Gaussian distribution is derived, The relationship holds.
here,
Here, T is a standardized transformation matrix defined by HM.

[A] _u , _u are the (u, u) -th component of the matrix A. The residual is modeled as a Gaussian distribution as follows:
If N = 8,

Since the maximum value in the equation (11) is generated by the DC component,
It becomes.

When the distribution of the inter prediction residual distribution can be approximated by an expression using SAD, the relationship between the SAD and the transform coefficient threshold in the case of an 8 × 8 block is derived as an inequality.
If N = 8,
It becomes. Here, n is a parameter that controls the detection rate of AZB.

Derivation of AZB thresholds for other sizes can also be calculated as follows according to the same procedure.

Next, threshold value adjustment for different depths of blocks will be described. Among the above-described threshold values (the right side of the equations (15) to (18), hereinafter, residual threshold value), n is an important factor for controlling the AZB detection rate. By setting n to be small, transform quantization is skipped in more blocks, but on the other hand, more false detections may occur. Here, a technique for setting a more appropriate n for transform quantization of different sizes will be described. From FIG. 5, it can be seen that the processing of one TU is calculated in order from the largest size to the smallest size. The original process described in HM computes the overall candidate transform quantization and selects the optimal solution based on RD optimization criteria. This process can also be described in the frequency domain as shown in FIG. The three graphs shown in FIG. 6 show the energy after calculating the same TU with different transform sizes. From these comparisons, it can be seen that the generation ratio of AZB blocks is large in smaller size conversion. In order to correct the influence of termination with a large size, the above-described residual threshold is adjusted as follows.
Here, _CN is a constant corresponding to a different size. w _N is an adjustment factor shown in Table 3, which is one example obtained results of the experiment.

  Next, effects produced by this embodiment will be described. First, encoding efficiency will be described. In order to compare the coding efficiency, generally, an RD curve in which a generated bit amount (bitRate) and distortion (distortion) of an original image are two-dimensionally graphed is used. For the purpose of avoiding the slight difference and obscure, the reference `` G. Bjontegard, '' Calculation of average PSNR differences between RD-curves, '' ITU-T VCEG-M33, 2001. The BDBR described in the above is adopted.

This is an average bit rate difference between the algorithm based on this embodiment and the performance of the conventional HM (hereinafter referred to as default HM), and measures the video quality as shown in Table 4. It is. + In BDBR represents a bit rate gain. HEVC's new coding tool rate distortion optimization based RDOQ and non-square quadtree transform (NSQT) are both combined and evaluated in a random access setup. The results are classified into a first method I with RDO and NSQT turned off, a second method II with RDOQ turned on, and a third method III with NSQT turned on.

Table 5 shows the reduction in transform quantization time. Here, the time saving of transform quantization is calculated based on the equation (19). TIME _HM is the transform quantization time consumed by HM6.3. TIME _proposal is the transform quantization time of the algorithm described above. It achieved a time reduction of 51% to 42% with RDOQ and NSQT. The algorithm described above is described in H.C. Compared to H.264 conventional methods, it plays a more important role not only in transform quantization but also in saving total encoding time. By adding a quadtree structure for traversal computation and large size transform quantization, the algorithm described above has the effect of accelerating the entire encoder, using the AZB method.
Here, TS is the time of conservative conversion quantization as a percentage between the above-described method and HEVC.

Next, the detection rate of AZB will be described. In order to compare the AZB detection accuracy, the detection rates of the two test class sequences are shown in FIG. Here, the horizontal axis represents the QP value, and the vertical axis represents the AZB detection rate (DR). Calculated as follows for AZB DR.

  As described above, detection of all-zero blocks of various block sizes can be derived by Gaussian modeling of residuals and transformed coefficients. Also, in order to overcome the influence of large block size determination, the weight adjustment can be performed by analyzing the conversion frequency of the corresponding magnitude of the residual.

  The transform quantization apparatus in the above-described embodiment may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included and a program held for a certain period of time. Further, the program may be for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in the computer system. It may be realized using hardware such as PLD (Programmable Logic Device) or FPGA (Field Programmable Gate Array).

  As mentioned above, although embodiment of this invention has been described with reference to drawings, the said embodiment is only the illustration of this invention, and it is clear that this invention is not limited to the said embodiment. is there. Therefore, additions, omissions, substitutions, and other modifications of the components may be made without departing from the technical idea and scope of the present invention.

  In addition to being able to detect AZB for various block sizes, it is also possible to apply applications where it is essential to enable adjustment by weighting for different size transform units.

  DESCRIPTION OF SYMBOLS 1 ... Inter prediction part, 2 ... AZB determination part, 3 ... Integer transform quantization part, 4 ... Entropy encoding part

Claims (4)

A transform quantization method for calculating a quantized transform coefficient from an inter prediction residual in image coding and a sum of absolute differences thereof,
Transformation of N × N size block (N = 2 ⁿ , where n is an integer of 3 or more) that is the object of image coding from a transformation coefficient threshold table showing the relationship between transformation quantization size, QP value, bit depth and transformation coefficient threshold A transform coefficient threshold specifying step for obtaining a transform coefficient threshold corresponding to the quantization size N, the input QP value, and the bit depth;
A difference absolute value sum upper limit calculation step in which a result obtained by multiplying the transform coefficient threshold by an adjustment coefficient determined for the transform quantization size is a difference absolute value sum upper limit value;
An AZB determination step of comparing the difference absolute value sum and the difference absolute value sum upper limit value;
When the difference absolute value sum is determined to be smaller than the difference absolute value sum upper limit value by the AZB determination step, a quantization transform coefficient determination step that sets the quantization transform coefficient to zero;
If it is determined in the AZB determination step that the difference absolute value sum is greater than or equal to the difference absolute value sum upper limit value, the block of N × N size is divided into four (N / 2) × (N / And 2) a division step of dividing the block into sizes and executing the AZB determination step again.   The transform coefficient threshold table is derived when the inter prediction residual and the transform coefficient follow a Gaussian distribution, and all the quantized transform coefficients of the block to be image-coded are all smaller than 1. 2. The transform quantization method according to claim 1, wherein the value is defined as a transform coefficient threshold value. A transform quantization apparatus for calculating a quantized transform coefficient from an inter prediction residual in image coding and a sum of absolute differences thereof,
Transformation of N × N size block (N = 2 ⁿ , where n is an integer of 3 or more) that is the object of image coding from a transformation coefficient threshold table showing the relationship between transformation quantization size, QP value, bit depth and transformation coefficient threshold Transform coefficient threshold specifying means for obtaining a transform coefficient threshold corresponding to the quantization size N, the input QP value, and the bit depth;
A difference absolute value sum upper limit calculating means that sets a result obtained by multiplying the transform coefficient threshold by an adjustment coefficient determined for the transform quantization size as a difference absolute value sum upper limit value;
AZB determination means for comparing the difference absolute value sum and the difference absolute value sum upper limit value;
When the difference absolute value sum is determined to be smaller than the difference absolute value sum upper limit value by the AZB determination unit, a quantization transform coefficient determination unit that sets the quantization transform coefficient to zero,
If it is determined by the AZB determination means that the sum of absolute differences is greater than or equal to the upper limit of absolute differences, the four blocks of (N / 2) × (N / 2) Dividing means that divides the block into size blocks and executes the determination by the AZB determining means again.   A transform quantization program for causing a computer to execute the transform quantization method according to claim 1.