VVC rapid CU division method based on random forest and residual network

Abstract

The invention relates to a VVC fast CU division method based on integrated learning, which belongs to the field of video coding and comprises the following steps: the method is mainly divided into a first stage dominated by a random forest classifier and a second stage dominated by a residual network, in the first stage, the random forest classifier is used for predicting the division state of a coding tree unit, and in the second stage, the residual network is used for predicting the division state of the coding tree unit; and the output result is either segmented or not segmented. In the second stage, the pixel brightness value of a 32 * 32 coding unit is input into the trained residual network classifier. After the residual network is calculated, the sub-coding units are classified into one of 17 category labels, and each label represents a 32 * 32 coding unit division structure. The method can be used for coding under VVC intra-frame coding configuration, the coding time of VVC is remarkably shortened under the condition that the coding efficiency and the coding quality are hardly reduced, and the method can be applied to video conferences and other application scenes with high requirements for coding real-time performance.

Description

Technical Field

[0001] The present invention belongs to the field of video coding, and more specifically, provides a CU fast partitioning method suitable for universal video coding VVC (Versatile Video Coding), which can be applied to video coding scenarios with high real-time requirements such as video conferencing and surveillance video. Background Art

[0002] In the current research process, the coding unit (CU) division suitable for VVC is extremely complex. If you rely solely on the bag tree model or the residual network to design a fast division algorithm, the CU size factor must be taken into consideration. In this way, the difficulty of designing the classifier model will rise sharply. The reason is that with the reduction of the CU size, the prediction accuracy of these two types of classifiers will be greatly reduced. When the CU is 64×64, they can still maintain a high prediction accuracy; but once the CU is reduced to 16×16, the prediction effect is far from satisfactory. From the perspective of information theory, the amount of information carried by smaller CUs is quite limited, which undoubtedly increases the uncertainty of the prediction results.

[0003] In order to resolve the rate-distortion loss problem caused by the low prediction accuracy of the classifier due to the small size of the CU, the algorithm needs to regard CU division as a three-classification task. Although it is expected to find a balance between rate distortion and encoding time by enabling a three-output classifier, this often requires the use of complicated threshold calculation methods similar to Algorithm 2 to accurately set high and low thresholds, which is difficult to implement in practice. In view of the fact that deep learning technology has more advantages than traditional machine learning in terms of prediction accuracy, most fast division algorithms resolutely abandon the three-output strategy. Although deep learning methods have made CU fast division no longer limited to three-output classifiers, it still needs to be judged layer by layer along the deep structure of the CU. As a result, there is a thorny problem that the classifier's prediction performance will gradually decline as the CU depth increases.

[0004] The key to overcoming the above-mentioned challenges is to devise an end-to-end coding tree unit (CU) partition structure decision algorithm, thereby fully tapping the powerful computing and classification potential of neural networks and avoiding the tedious process of repeatedly iterating segmentation and predicting CUs at a deep level. However, there are as many as 83,522 combinations of coding tree unit partition structures, and even neural networks have a hard time processing such a large number of classifications. Fortunately, there are only 17 CU partition structures of 32×32 size, which undoubtedly opens up a feasible path for the design of end-to-end algorithms.

[0005] This method innovatively combines random forest and residual network technology to overcome the problem of fast determination of CU partition structure in VVC intra-frame coding, and then proposes an end-to-end solution. This method first uses a two-output random forest classifier to predict the segmentation of CU; then, it transforms the difficulty of determining the partition structure of CU at depth 1 into a multi-classification task with 17 categories, and then processes it by building and training an adaptive residual network model, and finally realizes the rapid determination of the CU partition structure. Summary of the invention

[0006] The present invention aims to solve the above problems of the prior art. A VVC fast CU division method based on random forest and residual network is proposed. The technical solution of the present invention is as follows:

[0007] S1, obtain the coding information of the current coding unit CU to be divided in the coding frame, filter out the coding unit with a size of 64×64, and enter step S2;

[0008] S2, loading the pre-trained random forest classifier to predict the segmentation state of the coding tree unit, the output result either indicates segmentation or not segmentation, if not segmentation, proceed to step S5, if segmentation, proceed to step S3;

[0009] S3. If the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit is divided into four 32×32 sub-coding units, and the process proceeds to step S4.

[0010] S4. Load the pre-trained residual network classifier to determine whether the 32×32 sub-coding unit is segmented. If segmented, proceed to step S5.

[0011] S5. Enter the coding tree unit division structure determination module, and the encoder performs encoding operations according to the finally determined division structure.

[0012] S6. End the current CU division process.

[0013] Furthermore, in the step S1, CUs with a coding unit size of 64×64 are screened out because the method is mainly used for dividing CUs with a size of 64×64.

[0014] Furthermore, in step S2, the random forest classifier selects the top 10 candidate features as the activation features of the random forest classifier. And 20 different random forest classifiers are trained for 4 quantization parameter values ​​and 5 video resolutions in the standard test sequence. If the prediction result is no segmentation, the algorithm will directly enter the coding tree unit division structure determination process, and the encoder will fully encode the 64×64 coding tree unit without any segmentation.

[0015] Furthermore, in step S3, if the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit will be divided into four 32×32 sub-coding units. These four sub-coding units are then passed to the residual network classifier in turn for further prediction processing. At this time, VVC will skip the rate-distortion optimization process for the current coding tree unit.

[0016] Furthermore, in step S4, the pixel brightness values ​​of the 32×32 coding units will be input into the trained residual network classifier. After calculation, the residual network classifies these sub-coding units into one of 17 category labels, each label representing a 32×32 coding unit division structure. For the coding tree units that are judged as split in the random forest classifier, their corresponding four sub-coding units will be distinguished by the residual network classifier and obtain their respective category labels. The residual network classifier extracts and classifies the input coding units through multi-layer convolution and nonlinear activation functions. Assuming that the residual network consists of l layers, for the input x, the forward propagation process of the residual network can be expressed as:

[0017] y l =σ(W lyl-1 +b l )+y l-1 , l = 1, ..., L (1)

[0018] Among them, y l is the output of layer l, W lyl-1 and b l are the weight and bias of the lth layer, respectively, and is a nonlinear activation function (such as ReLU).

[0019]

[0020] in, The output of the last layer is normalized to represent the category probability distribution, and the argmax operation selects the category with the largest probability as the output.

[0021] For a 64×64 coding unit x, its output after passing through the residual network is:

[0022] R(x) = argmax(softmax(y L )) (3)

[0023] Furthermore, in step S5, entering the coding tree unit division structure determination module means that the category labels of the four sub-coding units, together with the segmentation labels of the coding tree units, are completely input into the coding tree unit division structure determination module. The encoder will strictly carry out subsequent encoding operations in an orderly manner according to the finalized division structure to ensure that the encoding results meet the expected standards.

[0024] The advantages and beneficial effects of the present invention are as follows:

[0025] Aiming at the problem that the latest international video coding standard H.266 / VVC has high computational complexity and long encoding time, the present invention proposes a CU fast partitioning method based on ensemble learning. First, in order to adapt to the VVC, the coding tree unit (CU) partitioning structure fast determination algorithm integrating random forest and residual network is mainly divided into two sections: one is the first stage based on random forest classifier; the other is the second stage based on residual network. In the first stage of the algorithm, the random forest classifier fully predicts the current segmentation situation of the coding tree unit, and its output result is either one or the other, either it is determined to be segmented or it is determined to maintain the original state and not segmented. Entering the second stage, the pixel brightness value corresponding to the coding unit of size 32×32 will be accurately input into the pre-trained residual network classifier. Residual network relies on its powerful computing power, and after a series of complex calculations, it is fully input into the coding tree unit partitioning structure determination module. The encoder will strictly carry out subsequent encoding operations in an orderly manner according to the finalized partitioning structure to ensure that the encoding results meet the expected standards. The present invention can be used for video coding under VVC intra-frame coding configuration, significantly reducing the VVC coding time without almost reducing the coding efficiency and coding quality, and can be applied to application scenarios such as video surveillance that have high requirements for real-time coding.

[0026] The innovation of the present invention mainly lies in claim 3, claim 4 and claim 5. The two classifiers contained in claim 3 and claim 5 are the core of the present invention, and they divide the algorithm into two stages. In the first stage of the algorithm, the random forest classifier predicts the current segmentation status of the coding tree unit, and its output result is either one or the other, either it is judged to be segmented, or it is judged to maintain the original state and not segmented. If the random forest classifier determines that the coding tree unit needs to be segmented, then the coding tree unit will be quickly disassembled into 4 32×32 sub-coding units. Subsequently, these 4 sub-coding units will be sequentially transmitted to the residual network classifier in the second stage for further depth prediction and fine processing. Entering the second stage, the pixel brightness values ​​corresponding to the 32×32 coding units will be accurately input into the pre-trained residual network classifier. Relying on its powerful computing power, the residual network will accurately classify these sub-coding units into one of the 17 category labels after a series of complex calculations. Each category label here is of special significance, and each corresponds to a 32×32 coding unit division structure. For the coding tree units that are determined to be segmented in the first stage, the four sub-coding units under it will be judged by the residual network classifier one by one, and then obtain exclusive category labels. Finally, the category labels of these four sub-coding units, together with the segmentation labels of the coding tree units, are fully input into the coding tree unit division structure determination module. The advantages of random forests and residual networks are fully utilized to achieve rapid CU division judgment, so the method in the present invention is highly innovative. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a VVC fast CU division method based on random forest and residual network provided by the present invention;

[0028] Figure 2 A VVC fast CU division method based on random forest and residual network provided by the present invention is implemented in the structure diagram;

[0029] Figure 3 is the relationship between categories and structures in the second stage of the algorithm; DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0031] The technical solution of the present invention to solve the above technical problems is:

[0032] Figure 1 The present invention provides a VVC fast CU division method based on random forest and residual network, and the method of the present invention comprises the following steps:

[0033] S1, obtain the coding information of the current coding unit CU to be divided in the coding frame, filter out the coding unit with a size of 64×64, and enter step S2;

[0034] S2, loading the pre-trained random forest classifier to predict the segmentation state of the coding tree unit, the output result either indicates segmentation or not segmentation, if not segmentation, proceed to step S5, if segmentation, proceed to step S3;

[0035] S3. If the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit is divided into four 32×32 sub-coding units, and the process proceeds to step S4.

[0036] S4. Load the pre-trained residual network classifier to determine whether the 32×32 sub-coding unit is segmented. If it is segmented, proceed to step S5. If it is not segmented, proceed to step S3.

[0037] S5. Enter the coding tree unit division structure determination module, and the encoder performs encoding operations according to the finally determined division structure.

[0038] S6. End the current CU division process.

[0039] Preferably, in step S1, CUs with coding units of 64×64 size are screened out because the method is mainly for dividing CUs with 64×64 size.

[0040] Preferably, in step S2, the random forest classifier selects the top 10 most important candidate features as activation features of the random forest classifier. And 20 different random forest classifiers are trained for 4 quantization parameter values ​​and 5 video resolutions in the standard test sequence. If the prediction result is no segmentation, the algorithm will directly enter the coding tree unit division structure determination process, and the encoder will fully encode the 64×64 coding tree unit without any segmentation.

[0041] Preferably, in step S3, if the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit will be divided into 4 32×32 sub-coding units. These 4 sub-coding units are then passed to the residual network classifier in turn for further prediction processing. At this time, VVC will skip the rate-distortion optimization process for the current coding tree unit.

[0042] Preferably, in step S4, the pixel brightness value of the 32×32 coding unit will be input into the trained residual network classifier. After calculation, the residual network classifies these sub-coding units into one of 17 category labels, each label representing a 32×32 coding unit division structure. For the coding tree unit that is judged as a segmentation in the random forest classifier, its corresponding four sub-coding units will be distinguished by the residual network classifier respectively and obtain their respective category labels. The residual network classifier extracts and classifies the input coding units through multi-layer convolution and nonlinear activation functions. Assuming that the residual network consists of l layers, for the input x, the forward propagation process of the residual network can be expressed as:

[0043] y l =σ(W lyl-1 +b l )+y l-1 , l = 1, ..., L (1)

[0044] Among them, y l is the output of layer l, W lyl-1 and b l are the weight and bias of the lth layer, respectively, and is a nonlinear activation function (such as ReLU).

[0045]

[0046] in, The output of the last layer is normalized to represent the category probability distribution, and the argmax operation selects the category with the largest probability as the output.

[0047] For a 64×64 coding unit x, its output after the residual network is:

[0048] R(x) = argmax(softmax(y L )) (3)

[0049] Preferably, in step S5, entering the coding tree unit division structure determination module means that the category labels of the four sub-coding units, together with the segmentation labels of the coding tree units, are completely input into the coding tree unit division structure determination module. The encoder will strictly carry out subsequent encoding operations in an orderly manner according to the finalized division structure to ensure that the encoding results meet the expected standards.

[0050] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0051] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A VVC fast CU division method based on random forest and residual network, characterized in that: The following steps are involved: S1, obtain the coding information of the current coding unit CU to be divided in the coding frame, filter out the coding unit with a size of 64×64, and enter step S2; S2, loading the pre-trained random forest classifier to predict the segmentation state of the coding tree unit, the output result either indicates segmentation or not segmentation, if not segmentation, proceed to step S5, if segmentation, proceed to step S3; S3. If the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit is divided into four 32×32 sub-coding units, and the process proceeds to step S4. S4. Load the pre-trained residual network classifier to determine whether the 32×32 sub-coding unit is segmented. If segmented, proceed to step S5. S5. Enter the coding tree unit division structure determination module, and the encoder performs encoding operations according to the finally determined division structure. S6. End the current CU division process.

2. A VVC fast CU division method based on random forest and residual network according to claim 1, characterized in that In the step S1, CUs with a coding unit size of 64×64 are selected because the method is mainly for dividing CUs with a size of 64×64.

3. A VVC fast CU division method based on random forest and residual network according to claim 1, characterized in that In step S2, the random forest classifier selects the top 10 candidate features as the activation features of the random forest classifier. And 20 different random forest classifiers are trained for 4 quantization parameter values ​​and 5 video resolutions in the standard test sequence. If the prediction result is no segmentation, the algorithm will directly enter the coding tree unit division structure determination process, and the encoder will fully encode the 64×64 coding tree unit without any segmentation.

4. The VVC fast CU division method based on random forest and residual network according to claim 1 is characterized in that In step S3, if the random forest classifier determines that the coding tree unit needs to be split, the coding tree unit will be divided into 4 32×32 sub-coding units. These 4 sub-coding units are then passed to the residual network classifier in turn for further prediction processing. At this time, VVC will skip the rate-distortion optimization process for the current coding tree unit.

5. The VVC fast CU division method based on random forest and residual network according to claim 1 is characterized in that In step S4, the pixel brightness values ​​of the 32×32 coding unit will be input into the trained residual network classifier. After calculation, the residual network classifies these sub-coding units into one of 17 category labels, each label representing a 32×32 coding unit division structure. For the coding tree unit that is judged as a segmentation in the random forest classifier, its corresponding four sub-coding units will be distinguished by the residual network classifier respectively and obtain their respective category labels. The residual network classifier extracts and classifies the input coding units through multi-layer convolution and nonlinear activation functions. Assuming that the residual network consists of l layers, for the input x, the forward propagation process of the residual network can be expressed as: y l =σ(W lyl-1 +b l )+y l-1 ,l=1,…,L (1) Among them, y l is the output of layer l, W lyl-1 and b l are the weight and bias of the lth layer, respectively, and is a nonlinear activation function (such as ReLU). in, The output of the last layer is normalized to represent the category probability distribution, and the argmax operation selects the category with the largest probability as the output. For a 64×64 coding unit x, its output after the residual network is: R(x)=argmax(softmax(y L )) (3)。 6. The VVC fast CU division method based on random forest and residual network according to claim 1 is characterized in that In step S5, entering the coding tree unit division structure determination module means that the category labels of the four sub-coding units, together with the segmentation labels of the coding tree units, are completely input into the coding tree unit division structure determination module. The encoder will strictly carry out subsequent encoding operations in an orderly manner according to the finalized division structure to ensure that the encoding results meet the expected standards.

Images