VVC all-zero block skipping method based on multi-level architecture

Abstract

The invention discloses a VVC all-zero block skipping method based on a multi-level framework, which comprises the following steps of: reading a transformation coefficient from a transformation block TU, comparing the amplitude of the maximum transformation coefficient with a pre-calculated threshold value, if the amplitude is smaller than the threshold value, judging that a true all-zero block GAZB is generated, skipping a quantization and coding process, otherwise, judging that the GAZB is not generated; a non-GAZB skipping cost is calculated through a distortion optimization algorithm, whether a non-GAZB coding process is skipped or not is judged, if the normal coding cost is larger than the non-GAZB skipping cost, the block is judged to be a pseudo all-zero block PAZB for skipping, and otherwise, the block is judged to be a non-all-zero block NAZB; and in the calculation process of the dependent quantization DQ, all-zero detection is carried out by counting context features of the CG so as to skip the coding process in which the CG is detected to be all-zero. According to the method, a machine learning technology is not used, the calculation complexity is reduced under the condition that the coding quality is ensured, and the coding speed is increased.

Description

Technical Field

[0001] The invention belongs to the field of video coding and relates to a VVC all-zero block skipping method based on a multi-level architecture. Background Art

[0002] The existing paper discloses the research on VVC inter-frame and quantization fast algorithm based on multi-level decision, which is used to optimize the inter-frame prediction module and quantization module in the VVC (Versatile Video Coding) video encoder to reduce the complexity of the encoder.

[0003] The algorithm optimizes the inter-frame prediction module and proposes a fast affine motion estimation (AME) algorithm based on multi-level decision-making. The complexity of the AME algorithm is reduced through three-level decision-making, including: Block partitioning: In the first level decision, the impact of the optimal inter-frame mode of the parent coding unit (CU) on the inter-frame mode of the sub-CU is analyzed through block partitioning, and the inter-frame mode of the sub-CU is predicted.

[0004] Affine motion search acceleration: In the second-level decision, the AMS process is accelerated through three steps, including skipping AMS early, terminating the iteration process early based on a custom value, and accelerating the fine-grained CPMV search.

[0005] Decision tree model prediction: In the third-level decision, 8 features are extracted and the optimal inter-frame mode of the CU is predicted using the offline trained decision tree model.

[0006] For the optimization of the quantization module, a fast quantization algorithm based on multi-level decision-making is proposed to perform a comprehensive detection of the all-zero block (AZB) to skip the quantization module, including: True all-zero block (GAZB) detection: In the first-level decision, the threshold is derived based on the hard decision quantization (HDQ) formula to detect true all-zero blocks.

[0007] Pseudo All-Zero Block (PAZB) Detection: In the second-level decision, Pseudo All-Zero Blocks are detected by analyzing the transform coefficient positions and proposing an adaptive threshold related to the quantization parameter (QP) and transform block (TU) size.

[0008] Prediction based on machine learning: In the third-level decision, eight influencing factors affecting the TU quantization results are extracted, and the offline trained fully connected neural network (FCNN) is used to predict the TU quantization results.

[0009] This technology uses a fully connected neural network FCNN when processing PAZB detection, with the aim of improving the detection accuracy of PAZB. Although this method improves the detection accuracy, FCNN has a total of 2 hidden layers, 1 input layer and 1 output layer in terms of the number of network layers and the number of nodes in each layer. This structure will increase the significant computational burden in practical applications, especially in the dependency quantization stage.

[0010] The traditional all-zero block detection method cannot be directly applied to the VVC coding standard because VVC introduces a new block division method and more complex coding tools, which makes the original detection method invalid. This technology breaks through the dilemma that the traditional all-zero block judgment method cannot be directly applied to VVC, but it still needs to be further optimized to establish an approximate bit rate estimation model and use the rate distortion optimization (RDO) method to judge PAZB. However, the bit rate model was originally used to estimate the actual bit rate value, and the bit rate value in the dependency quantization (DQ) actually represents the bit rate change value compared with the all-zero block. Therefore, the accuracy of the model is not high enough, which will cause errors in the judgment results of some TUs, and there is still room for improvement in the selection of feature variables. It introduces machine learning technology in TU-level feature analysis, extracts 8 influencing factors that affect the TU quantization results, and predicts based on FCNN. Although this method can optimize the prediction of TU quantization results to achieve more efficient video encoding, the introduction of machine learning models brings about the problem of high computational complexity, resulting in delays, affecting the real-time encoding, and the performance of the trained model on different video sequences may vary greatly. Here, the above method is still for TU-level all-zero block judgment, and all-zero transform coefficients (CG) will appear in dependency quantization, so there is still room for optimization in this part. Summary of the invention

[0011] In order to solve the above technical problems existing in the prior art, the present invention proposes a VVC all-zero block skipping method based on a multi-level architecture, avoids the use of machine learning, adopts an all-zero CG skipping strategy, and reduces the calculation steps by multi-level optimization of the quantization process, saving coding resources and reducing the calculation burden. The specific technical solution is as follows: A VVC all-zero block skipping method based on a multi-level architecture, comprising: Step 1: After reading the transform coefficient from the transform block TU, compare the amplitude of its maximum transform coefficient with the pre-calculated threshold. If it is less than the threshold, it is judged as a true all-zero block GAZB, and the quantization and encoding process can be skipped. Otherwise, it is non-GAZB and proceeds to the next step; Step 2: Build a bit rate estimation model and use the rate distortion optimization algorithm to calculate the cost of non-GAZB skipping, and compare it with the cost of normal encoding to determine whether to skip the non-GAZB encoding process. If the cost of normal encoding is greater than the cost of non-GAZB skipping, the block is judged as a pseudo all-zero block PAZB and skipped. Otherwise, it is judged as a non-all-zero block NAZB. Step 3: During the calculation of dependency quantization DQ, all-zero detection is performed by statistically analyzing CG context features to skip the encoding process of detecting all-zero CG.

[0012] Furthermore, the step 1 specifically includes: In VCC, DQ is described as: , in, To quantize the step, is the dead zone offset, Indicates a round-down operation. and are the transformation coefficients before and after scaling, >> is the shift operation on the variable, ; is the number of bits of the quantization step, bits Refers to the number of shifts; MF is a fixed scaling factor; For the true all-zero block GAZB, let the maximum scalar quantization value be less than an empirical coefficient ,Right now: , in, is the empirical coefficient, obtained by experiment; Combining the above inequality with the equation, we can get the threshold value after processing: , Further transformed into the following form: , Therefore, when the maximum transform coefficient amplitude in the TU is less than the threshold , then the TU is judged as a true all-zero block GAZB.

[0013] Furthermore, in step 2, the rate-distortion optimization algorithm is expressed as: , Where D represents distortion, R represents bit rate, and λ is the Lagrangian factor; When non-GAZB is judged as PAZB, it satisfies: , otherwise it is a non-all-zero block NAZB; This inequality And the equation By combining them, we get the following system of equations: , is the difference between the distortion when the block is encoded as an all-zero block and the distortion when the block is encoded as a non-all-zero block, The difference between the code rate generated when the block is encoded as a non-all-zero block and the code rate generated when the block is encoded as an all-zero block.

[0014] Further, in step 2, the bit rate estimation model is used to obtain , the specific expression is: , in All obtained by linear regression; Represents the number of non-zero coefficients in TU. It represents the sum of the absolute values ​​of the non-zero coefficients in TU; is the characteristic variable, which is the sum of all coefficients whose variable quantization coefficients are greater than 7.

[0015] Further, in step 2, the difference between the distortion when encoding the non-GAZB block as an all-zero block and the distortion when encoding the block as a non-all-zero block is , the specific expression is: , , Subtracting the two formulas, we get The calculation formula is: , Among them, M and N represent the width and height of the TU block respectively; represents the original transform coefficient, k is the coefficient index; The reconstructed value of a block encoded as PAZB after dequantization is 0, so the distortion formula of PAZB only contains the real transform coefficient. The transform coefficients of the block encoded as NAZB before the last non-zero coefficient position L are normally quantized, so the reconstructed value after dequantization is Still retained, the transform coefficients after the last non-zero coefficient position L will be quantized to 0, so only the real transform coefficients are retained.

[0016] Furthermore, the step 3 is specifically as follows: in the dependency quantization DQ calculation process in VVC, all-zero detection is performed by statistically analyzing the CG context features to screen out candidate CGs, and then the decision conditions are derived through the RDO formula, and then the approximate distortion accumulation model is used to judge the candidate CGs through the decision conditions. If the CG is determined to be an all-zero CG, the encoding process of the CG is skipped.

[0017] Furthermore, the decision condition derived from the rate-distortion optimization RDO formula is specifically: when calculating the rate-distortion cost generated by skipping a group of transform coefficients CG encoding Less than the rate-distortion cost of encoding the CG When, that is: , then the CG is encoded as all-zero CG.

[0018] Furthermore, the expression of the rate-distortion cost generated by skipping a group of transform coefficients CG encoding is: , when the CG encoding is skipped, the rate distortion cost is only the bit rate generated by encoding the all-zero CG flag; The expression of the rate distortion cost generated by encoding a set of transform coefficients CG is: , the CG is regarded as a non-zero CG, then its rate-distortion cost includes the coding rate and the reconstruction value of the coding and the distortion change value generated when the coefficient is coded as 0, where: , that is, the bit rate is always greater than 0. After removing the bit rate part, we get the inequality: , From this inequality, we can know that if the accumulated distortion change value of each CG coefficient is greater than the bit rate generated by skipping the CG encoding, then the CG should be skipped.

[0019] Furthermore, the approximate distortion accumulation model is expressed as: , That is, the smallest distortion among the candidate CGs is selected as the distortion of the optimal path, and accumulated, and then judged by the decision conditions derived from the rate-distortion optimization RDO formula to obtain a full zero CG.

[0020] Advantages of the present invention include: First, in PAZB detection, machine learning technology is not used. Instead, a TU-level bit rate estimation model is established to judge PAZB through RDO rate distortion optimization, thereby reducing the amount of calculation in the quantization and encoding process. Although this method may be slightly inferior to the machine learning method in detection accuracy, it has obvious advantages in computational complexity and can meet the needs of real-time encoding. Experimental results show that the algorithm of the present invention can significantly shorten the encoding time and keep the video quality almost unchanged.

[0021] Second, an all-zero CG skip strategy is adopted, which is an optimization method for all-zero block processing in the DQ decision process. By analyzing the quantization characteristics of transform coefficients and combining a simplified distortion accumulation model, all-zero CG is detected and skipped, thereby avoiding the high computational complexity brought by machine learning and speeding up the encoding speed. Although the encoding of some blocks is skipped, it may cause a small loss of quality, but by reasonably selecting the skip conditions, the encoding quality loss is ensured to be within an acceptable range, especially with the assistance of RDO rate-distortion optimization, skipping encoding has little impact on the final video quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of a VVC all-zero block skipping method based on a multi-level architecture of the present invention; Figure 2 It is a schematic diagram for explaining the scanning channel; Figure 3a~Figure 3d is a graph showing the relationship between the bit rate model proposed by the present invention and the actual bit rate value in the reference code when the QP settings are 22, 27, 32 and 37 respectively under the BasketballDrill test sequence of this embodiment; Figure 4 This is a schematic diagram of a context model for generating each coefficient of CG in this embodiment; Figure 5 It is the probability map of the all-zero CG when all the context elements of the CG of this embodiment are 0 and the QP settings are 22, 27, 32 and 37 respectively. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and technical effect of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0024] like Figure 1 As shown, the present invention discloses a VVC all-zero block skipping method based on a multi-level architecture, and the process is as follows: Step 1, detect GAZB: After reading the transform coefficient from the transform block TU, compare the amplitude of its maximum transform coefficient with the pre-calculated threshold. If it is less than the threshold, it is judged as GAZB and the quantization and encoding process can be skipped. Otherwise, proceed to the next step.

[0025] Step 2, detect PAZB: For non-GAZB, use the rate-distortion optimization algorithm to calculate the cost of skipping the block and compare it with the cost of normal encoding to determine whether to skip the encoding of the block.

[0026] Feature variables are extracted from non-GAZB and input into the bit rate estimation model constructed by the present invention to obtain the bit rate, and then the distortion generated when the TU is not skipped is calculated. If the bit rate consumed by encoding is greater than the distortion generated when the TU is skipped, the block is judged as PAZB, otherwise it is judged as a non-all-zero block NAZB.

[0027] Step 3, detect all-zero CG: In the dependency quantization DQ calculation process in VVC, there are still all-zero CGs inside the non-all-zero block NAZB that can be skipped. Therefore, the present invention uses the feature points of the all-zero CG to perform all-zero detection by counting the CG context features. First, the CG that may be zero is screened out, and then the decision condition is derived through the RDO formula. The approximate distortion accumulation model is used to judge the CG through the decision condition. If the CG is determined to be an all-zero CG, the encoding process of the CG is skipped.

[0028] Specifically, the detection of GAZB in step 1 includes the following: Each coefficient must be scalar quantized to generate four non-zero candidate values ​​when performing DQ, and the Viterbi algorithm is used to select the best candidate value. However, if the actual quantization value of the transform coefficient is 0 after scalar quantization, the distortion caused by the four candidate values ​​will be greater than the actual quantization value. Therefore, all transform coefficients can be scanned in advance to determine whether the TU is AZB through certain conditions.

[0029] The traditional scalar quantization formula is as follows: , in is the quantization coefficient after scalar quantization, is the original transformation coefficient, To quantize the step, is the dead zone offset, Indicates a floor operation.

[0030] In the actual application of DQ, in order to avoid the computational complexity caused by floating-point numbers, the transform coefficients must be scaled before the calculation. The relationship between the actual formula in the code and the original scalar quantization formula is as follows: , in and are the transformation coefficients before and after scaling, >> is the shift operation on the variable, .

[0031] When the largest scalar quantization value among all transform coefficients in a TU is also less than 1, the TU will be encoded as an all-zero block because the scalar quantization value will be rounded down. This condition can be converted into the following formula: , However, since the DQ process will be processed by the rate-distortion optimization algorithm, the TU will not be encoded as an all-zero block only when all scalar quantization values ​​are less than 1. If the rate-distortion cost of encoding the TU block is greater than the rate-distortion cost of skipping encoding the TU block, the current block will still be encoded as AZB. Therefore, an empirical coefficient is used To replace the right side of the above inequality, we get the following new inequality: , in, Obtained by experiments to achieve a balance between performance and quality.

[0032] Combining this inequality with the equation, we can get the following formula after processing: , The above formula can be transformed into the following form: , Therefore, when the maximum transform coefficient amplitude in the TU is less than , the TU is judged as GAZB.

[0033] The detection of PAZB in step 2 includes the following: After the GAZB judgment, there is at least one transform coefficient in the remaining TU blocks with a large enough amplitude. However, non-GAZB does not mean that the TU block will be retained. DQ uses a rate-distortion optimization method to screen out the optimal path. After the rate-distortion calculation, if the rate-distortion cost of encoding the current TU block is greater than the rate-distortion cost of skipping the encoding of the TU, the TU block will still be judged as PAZB and skipped. Therefore, from the perspective of rate-distortion, we can find a way to calculate the rate cost and distortion cost of the current TU block in advance. The rate-distortion cost formula is as follows: , Where D represents distortion, R represents bit rate, and λ is the Lagrangian factor. If the block is judged as PAZB, it should satisfy the following inequality: , If the rate-distortion cost of skipping the block is smaller than the rate-distortion cost of encoding the block as a non-all-zero block, then the block should be skipped.

[0034] This inequality And the equation By combining them, we can get the following system of equations: , is the difference between the distortion when the block is encoded as an all-zero block and the distortion when the block is encoded as a non-all-zero block, is the difference between the code rate when the block is encoded as a non-all-zero block and the code rate when the block is encoded as an all-zero block. Therefore, the next focus is how to obtain and .

[0035] In the part of the bit rate estimation model, Sun et al. have proposed a solution, and the present invention has made some improvements on this basis. Specifically, the bit rate estimation model of the present invention also uses and These two characteristic variables, and Respectively represent Norm and The norm is consistent with the original model, and its calculation formula is as follows: , , in, Represents the non-zero coefficient in TU The number of Represents the sum of the absolute values ​​of the non-zero coefficients in TU.

[0036] In the original model, the third variable is , which is the bit rate generated when encoding the last non-zero coefficient position. However, the bit rate generated by this part accounts for a very small proportion of the entropy coding process after DQ is completed. Figure 2 It can be seen that in addition to the bit rate generated by encoding syntax elements such as sig, gt1, par, and gt3 in regular mode, a large part of the bit rate comes from coefficients with DQ quantization values ​​greater than 3. These parts with quantization values ​​greater than 3 will be represented by rem syntax elements in the entropy coding stage, and their coding mode is bypass. Therefore, using the relevant features of coefficients with DQ quantization values ​​greater than 3 to replace the bit rate generated by the last non-zero coefficient position can make the bit rate estimation model more accurate.

[0037] and 、 Both feature variables are extracted at the transform coefficient stage, so the conditions of the DQ quantization value are converted into the conditions of the transform coefficient. The DQ quantization coefficient and the scalar quantization coefficient are related by the following formula: , Combining the above formula with By combining, we can get It is only possible to encode the rem syntax element in the entropy coding stage when s_gt7 , that is, the sum of all coefficients whose variable quantization coefficients are greater than 7, is used as the feature variable. s_gt7 The calculation formula is as follows: , After completing the above calculations, the updated bit rate estimation model is as follows: , in , all obtained by linear regression.

[0038] Figure 3a~Figure 3d The figure shows the relationship between the bitrate estimation model proposed in the present invention and the actual bitrate value in the reference code when the QP settings are 22, 27, 32 and 37 respectively under the BasketballDrill test sequence. It can be seen from the figure that under different QP settings, the model proposed in the present invention has a strong linear relationship with the actual bitrate in the VTM calculation process, and the slope of the trend line is very close to 1. Therefore, the bitrate estimation model proposed in the present invention can estimate the bitrate at the TU level more accurately, and can be used in the PAZB judgment method proposed later.

[0039] After completing the establishment of the bit rate estimation model, the next thing to do is to obtain The distortion of PAZB and NAZB can be expressed by the following two formulas respectively: , , Where M and N represent the width and height of the TU block, respectively. Because the reconstructed value of the block encoded as PAZB is 0 after inverse quantization, the distortion formula of PAZB only contains the real transform coefficient.

[0040] The transform coefficients of the block encoded as NAZB before the last non-zero coefficient position L are all normally quantized, so the reconstructed value after dequantization is Still retained. The transform coefficients after the last non-zero coefficient position L are quantized to 0, so only the real transform coefficients are retained.

[0041] Subtracting the two formulas, we get The calculation formula is: , Thus, two important variables of this judgment method are obtained: and , and then based on the above conditions, it can be determined whether the block is PAZB.

[0042] The detection of all zero CG in step 3 includes the following contents: After determining GAZB and PAZB, there is a lot of encoding time left. However, in the DQ process inside the TU, there is also a skipping of all zero CGs. The detection and skipping of this part of CG will be a breakthrough in the new method of shortening DQ encoding time.

[0043] Since the internal process of DQ coding is unpredictable and has a strong contextual relationship, it is impossible to find a very accurate theoretical model to complete the skipping of all-zero CG. Therefore, the present invention starts from a statistical point of view, finds the characteristics of all-zero CG, and analyzes them to determine whether the CG is all zero.

[0044] The DQ process is based on the rate-distortion method, where there are four candidate values ​​for calculation, and the path with the lowest rate-distortion cost under the current coefficient is selected to be retained. The calculation of the bit rate is highly dependent on the context, so the features of the all-zero CG can be found from the context across CGs.

[0045] like Figure 4 As shown in the figure, the black part is the coefficient to be encoded, and the gray part is the reference coefficient. The context model includes sig, Abs1 and Abs. Among them, sig is the number of non-zero coefficients in the template area, Abs1 is the sum of the absolute values ​​of the coding levels of the coefficients in the template area, and Abs is the sum of the absolute values ​​of the reconstruction values ​​of the coefficients in the template area.

[0046] Figure 5 It shows the probability of the CG being an all-zero CG when all the elements of the context are 0 and the QP settings are 22, 27, 32, and 37. Therefore, this condition can be used to filter out possible all-zero CGs first.

[0047] The process of DQ establishing a grid is also based on rate-distortion optimization. If a CG is finally encoded as an all-zero CG, the conditions it should meet are as follows: , Only when a group of transform coefficients is skipped, the rate-distortion cost incurred by CG coding , which is less than the rate distortion cost of encoding the CG When , the CG is encoded as all zero CG. and The formulas are as follows: , , When the CG coding is skipped, its rate-distortion cost is only the bit rate generated by coding the all-zero CG flag. If the CG is regarded as a non-all-zero CG, its rate-distortion cost includes the coding bit rate and the coding reconstruction value and the distortion change value generated when the coefficient is coded as 0.

[0048] because , that is, the bit rate is always greater than 0, and it is impossible to make a very accurate coefficient-level prediction for it, so eliminating this part of the bit rate, we can get the following inequality: , From this inequality, we can know that if the accumulated distortion change value of each CG coefficient is greater than the bit rate generated by skipping the CG encoding, then the CG should be skipped. There are four candidate values ​​for distortion in total, and it is impossible to accurately determine which candidate value will eventually be on the optimal path, so the present invention uses the following formula for approximate calculation: , The smallest distortion among the four candidate values ​​is selected as the distortion of the optimal path, and accumulated. Then, the above conditions are used for judgment to obtain which CGs are all-zero CGs.

[0049] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A VVC all-zero block skipping method based on a multi-level architecture, characterized in that: include: Step 1: After reading the transform coefficient from the transform block TU, compare the amplitude of its maximum transform coefficient with the pre-calculated threshold. If it is less than the threshold, it is judged as a true all-zero block GAZB, and the quantization and encoding process is skipped. Otherwise, it is non-GAZB and proceeds to the next step; Step 2: Build a bit rate estimation model and use the rate distortion optimization algorithm to calculate the cost of non-GAZB skipping, and compare it with the cost of normal encoding to determine whether to skip the non-GAZB encoding process. If the cost of normal encoding is greater than the cost of non-GAZB skipping, the block is judged as a pseudo all-zero block PAZB and skipped. Otherwise, it is judged as a non-all-zero block NAZB. Step 3: During the calculation of dependency quantization DQ, all-zero detection is performed by statistically analyzing CG context features to skip the encoding process of detecting all-zero CG.

2. The VVC all-zero block skipping method according to claim 1, characterized in that: The step 1 specifically includes: In VCC, DQ is described as: ， in, To quantize the step, is the dead zone offset, Indicates a round-down operation. and are the transformation coefficients before and after scaling, >> is the shift operation on the variable, , is the number of bits of the quantization step, bits Refers to the number of shifts; MF is a fixed scaling factor; For the true all-zero block GAZB, let the maximum scalar quantization value be less than an empirical coefficient ,Right now: ， in, is the empirical coefficient, obtained by experiment; Combining the above inequality with the equation, we can get the threshold value after processing: ， Further transformed into the following form: ， Therefore, when the maximum transform coefficient amplitude in the TU is less than the threshold , then the TU is judged as a true all-zero block GAZB.

3. The VVC all-zero block skipping method according to claim 1, characterized in that: In step 2, the rate-distortion optimization algorithm is calculated as follows: ， in D Indicated as distortion, R Indicates the bit rate, λ is the Lagrangian factor; When non-GAZB is judged as PAZB, it satisfies: , otherwise it is a non-all-zero block NAZB; This inequality And the equation By combining them, we get the following system of equations: ， is the difference between the distortion when the block is encoded as an all-zero block and the distortion when the block is encoded as a non-all-zero block, The difference between the code rate generated when the block is encoded as a non-all-zero block and the code rate generated when the block is encoded as an all-zero block.

4. The VVC all-zero block skipping method according to claim 3, characterized in that: In step 2, the bit rate estimation model is used to obtain , the specific expression is: ， in All obtained by linear regression; Represents the number of non-zero coefficients in TU. It represents the sum of the absolute values ​​of the non-zero coefficients in TU; is the characteristic variable, which is the sum of all coefficients whose variable quantization coefficients are greater than 7.

5. The VVC all-zero block skipping method according to claim 3, characterized in that: In step 2, the difference between the distortion when encoding a non-GAZB block as an all-zero block and the distortion when encoding the block as a non-all-zero block , the specific expression is: ， ， Subtracting the two formulas, we get The calculation formula is: ， Among them, M and N represent the width and height of the TU block respectively; represents the original transform coefficient, k is the coefficient index; The reconstructed value of a block encoded as PAZB after dequantization is 0, so the distortion formula of PAZB only contains the real transform coefficient. The transform coefficients of the block encoded as NAZB before the last non-zero coefficient position L are normally quantized, so the reconstructed value after dequantization is Still retained, the transform coefficients after the last non-zero coefficient position L will be quantized to 0, so only the real transform coefficients are retained.

6. The VVC all-zero block skipping method according to claim 1, characterized in that: The step 3 is specifically as follows: in the dependency quantization DQ calculation process in VVC, all-zero detection is performed by statistically analyzing the CG context features to screen out candidate CGs, and then the decision conditions are derived through the RDO formula. Then, an approximate distortion accumulation model is used to judge the candidate CGs through the decision conditions. If the CG is determined to be an all-zero CG, the encoding process of the CG is skipped.

7. The VVC all-zero block skipping method according to claim 6, characterized in that: The decision condition derived from the rate-distortion optimization RDO formula is as follows: when calculating the rate-distortion cost generated by skipping a set of transform coefficients CG encoding Less than the rate-distortion cost of encoding the CG When, that is: , then the CG is encoded as all-zero CG.

8. The VVC all-zero block skipping method according to claim 7, characterized in that: The expression of the rate distortion cost generated by skipping a group of transform coefficients CG encoding is: , when the CG encoding is skipped, the rate distortion cost is only the bit rate generated by encoding the all-zero CG flag; The expression of the rate distortion cost generated by encoding a set of transform coefficients CG is: , the CG is regarded as a non-zero CG, then its rate-distortion cost includes the coding rate and the reconstruction value of the coding and the distortion change value generated when the coefficient is coded as 0, where: , that is, the bit rate is always greater than 0. After removing the bit rate part, we get the inequality: ， From this inequality, we can know that if the accumulated distortion change value of each CG coefficient is greater than the bit rate generated by skipping the CG encoding, then the CG will be skipped.

9. The VVC all-zero block skipping method according to claim 8, characterized in that: The approximate distortion accumulation model is expressed as: ， That is, the smallest distortion among the candidate CGs is selected as the distortion of the optimal path, and accumulated, and then judged by the decision conditions derived from the rate-distortion optimization RDO formula to obtain a full zero CG.

Images