CN104837019B - AVS to HEVC optimization video transcoding methods based on SVMs - Google Patents

Abstract

An optimized video transcoding method from AVS to HEVC based on support vector machine, by collecting the feature vector of AVS code stream, and using the support vector machine to learn it and obtain the training model, the extracted AVS feature vector is divided into HEVC The CU at the corresponding position in the CU can be divided or not divided into two categories. In the transcoding stage, the training model is used to predict whether the CU needs to be divided. When the current CU needs to be divided, the 2N×2N mode and the SKIP mode are calculated at the current HEVC depth. , and select the optimal prediction mode from these two modes. When it is predicted that the current CU does not need to be divided, the optimal mode selection is performed according to the HEVC standard encoding process. The present invention combines the basic idea of machine learning, divides the entire transcoding process into a training phase and a transcoding phase, obtains a training model through learning, predicts the division of CUs in HEVC, and combines the fast mode selection algorithm to improve the speed of transcoding , and ensure the overall video quality of the transcoded video.

Description

AVS to HEVC optimized video transcoding method based on support vector machine

technical field

The present invention relates to a technology in the field of video signal processing, in particular to an optimized video transcoding method from AVS to HEVC based on a support vector machine.

Background technique

Video transcoding technology is to decode and re-encode the compressed code stream to obtain the target code stream that meets the requirements. With the wide application and rapid development of multimedia technology and the Internet, the transmission of various video data on the network has become the development trend of network technology. At present, a variety of video coding standards have emerged, including MPEG-4, MPEG-2, H.264, AVS, HEVC, etc. Due to the variety of video resources, as well as the differences in the display capabilities, storage capabilities, and processing capabilities of different terminal devices, users have different requirements for video in different scenarios. Therefore, how to achieve efficient conversion Code, so that it can be adapted to different hardware devices and network transmission environments, has been widely concerned by the industry. HEVC is currently the latest video coding standard, and its compression efficiency is about 50% higher than that of H.264 and other standards. It will be more and more widely used. AVS is a standard independently developed by my country. It has the same coding performance and lower coding complexity as H.264, and has an important influence in the field of video applications. At present, there are many compressed AVS streams, which will coexist with HEVC and other standards for a long time. Therefore, the realization of video conversion between AVS and HEVC standards has become an important research direction.

Machine learning is used in many research fields, including artificial intelligence, machine translation, data mining, text recognition and commercial fields. With the development of digital multimedia technology and transmission network, it is also continuously applied to video search, video analysis and Research directions such as video encoding and transcoding, and the application of machine learning algorithms in the field of video transcoding is also increasing. Support vector machine is an important method of machine learning. It extracts decoding information from the front part of the input code stream, and uses support vector machine to learn the correspondence between AVS information and HEVC encoding mode. In the subsequent encoding process , directly predict the HEVC encoding mode based on the decoded AVS information, without the need for a complete iterative traversal process, so as to achieve the purpose of reducing the complexity of video conversion. Therefore, how to learn an accurate training model and achieve accurate prediction to reduce the complexity of video conversion and improve the rate of video transcoding by using the machine learning method of support vector machine has become an important topic of current research.

After searching the prior art, it was found that Chinese Patent Document No. CN104320667A was published (announced) on 2015.01.28, disclosing a multi-process optimized encoding system, including several parallel encoders, look-ahead buffers and secondary encoders. The input of the look-ahead buffer is connected to the output of the parallel encoder, the output of the look-ahead buffer is connected to the input of the secondary encoder, and its method is disclosed, including the first encoding stage, the optimization selection stage and the second There are 3 steps in the encoding stage. The first encoding stage is encoded by several parallel encoders at the same time. The look-ahead buffer optimizes the results obtained in the first encoding stage to obtain the optimal encoding path. The secondary encoder is optimized according to The optimal encoding path obtained in the selection stage is encoded for the second time to obtain the final and optimal encoding result. The technology has higher performance, quality, bandwidth efficiency, better encoding/transcoding results, is very easy to configure and is very flexible, and can be used for both 4K and UHD applications with high video quality, and mobile video applications with ultra-efficient bandwidth. However, the input of this technology is the original code stream, and the encoding information contained in the compressed code stream cannot be used. Moreover, for HEVC’s more complex encoders, the encoding paths are diverse, and the implementation complexity will be relatively high.

Contents of the invention

Aiming at the above-mentioned deficiencies in the prior art, the present invention proposes an optimized video transcoding method from AVS to HEVC based on a support vector machine, adopts a simplified video transcoding framework, and extracts the feature vector from the AVS code stream during the training phase The training model is obtained through training, and then the model is used in the transcoding stage to perform regional transcoding on the division of coding units in HEVC. At the same time, combined with the fast mode selection algorithm, the complexity of coding unit division and mode selection in the re-encoding phase is reduced, and the complexity of the mode selection is greatly improved. The speed of video transcoding is improved, and at the same time, the quality of transcoding video is limited.

The present invention is achieved through the following technical solutions:

The present invention collects the eigenvectors of the AVS code stream, uses the support vector machine to learn them and obtains the training model, and divides the extracted AVS eigenvectors into two categories: CU division or non-division at corresponding positions in HEVC. In the encoding phase, the training model is used to predict whether the CU needs to be divided.

The AVS feature vector is collected from the AVS code stream, including information such as macroblock coding mode, motion vector and transformation coefficient.

The macroblock coding mode refers to the mode information of each macroblock in the AVS.

The motion vector refers to: the average motion vector size of the macroblock in the AVS.

The transform coefficient refers to the number of non-zero coefficients in the discrete cosine transform (DCT) coefficients of AVS encoding.

The corresponding relationship refers to whether the AVS feature vector and the coding unit in HEVC need to be divided into a training model obtained through learning, that is, a mapping relationship.

For the above prediction, when the current CU needs to be divided, the 2N×2N mode and the SKIP mode are respectively calculated at the current HEVC depth, and the optimal prediction mode is selected from these two modes. When the current CU is predicted If division is not required, optimal mode selection is performed according to the HEVC standard encoding process.

technical effect

Compared with the prior art, the present invention combines the basic idea of machine learning and the learning method of support vector machine to obtain a training model for video learning and predict the HEVC encoding mode in the subsequent transcoding process, thereby reducing the complexity of transcoding and improving the transcoding speed . The entire transcoding process is divided into two stages, namely the training stage and the transcoding stage. In the training phase, the AVS feature vector and the CU division information of the corresponding position in HEVC are extracted, and the corresponding relationship between the two is obtained according to the support vector machine algorithm and tool training, that is, the training model; in the transcoding phase, the model obtained in the training phase is used, According to the feature vector of AVS, predict whether the CU at the corresponding position of HEVC is divided when the depth is 0 and 1. If it is predicted that the current CU needs to be divided, then only SKIP and 2Nx2N mode selection is performed for the CU of the current depth. If the current CU is predicted to be not If division is required, the optimal mode is selected according to the HEVC standard process. Through the combination of machine learning and fast mode selection methods, the corresponding training model can be adaptively obtained according to the characteristics of each sequence, which greatly reduces the computational complexity in the transcoding process while ensuring that the quality of the transcoded video is limited. Thus, the transcoding speed is improved and the transcoding time is saved.

Description of drawings

Fig. 1 is a frame diagram of the transcoding of the present invention;

Fig. 2 is a flowchart of the present invention.

detailed description

The embodiments of the present invention are described in detail below. This embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operating procedures are provided, but the protection scope of the present invention is not limited to the following implementation example.

Example 1

As shown in Figure 2, this embodiment is divided into the following four steps:

Step 1. Collecting the feature vectors of the AVS code stream, specifically: collecting the CU division information of corresponding positions in the AVS code stream and corresponding HEVC depths of 0 and 1.

The AVS feature vector is collected from the AVS code stream, including information such as macroblock coding mode, motion vector and transformation coefficient.

The macroblock coding mode refers to: when the depth of HEVC corresponding to AVS is 0, each CU contains 16 macroblocks, and therefore contains 16 features; when the depth is 1, each CU contains There are 4 macroblocks, corresponding to 4 mode features.

The motion vector refers to: the motion vector in the AVS is 8×8, and a macroblock contains 4 motion vector basic units, and each motion vector is respectively moduloed, and then the four motion vectors in a macroblock are The mean value of the modulus is used as a feature. Similarly, when the depth is 0, there are 16 motion features in the CU, and when the depth is 1, there are 4 motion features. The average value v _avg of the magnitude of the motion vector is equal to Where N is the number of 8×8 blocks included in the current coding unit, and v _x (i) and v _y (i) are the horizontal component and vertical component of the i-th 8×8 block motion vector in the coding unit, respectively.

The transformation coefficient refers to: DCT in AVS is also based on 8×8 blocks, and the number of non-zero coefficients in all DCT coefficients in a macro block is used as a feature. Therefore, when the depth is 0, there are 16 DCT coefficients Features, when the depth is 1, it contains 4 DCT coefficient features.

Step 2. Use the support vector machine to learn the AVS feature vectors obtained in step 1 and obtain a training model, which is used to classify the coding units when the HEVC depth is 0 and 1, and divide the extracted AVS feature vectors into two categories , that is, the CU division or no division at the corresponding position in HEVC.

Step 3. Use the training model obtained in step 2 to directly predict whether the current CU needs to be divided. When the prediction result is that no division is required, the division will be terminated and no further mode selection will be performed. Otherwise, it will enter the next step of HEVC. Iterate deeply and perform step 4.

Step 4: Perform 2N×2N mode and SKIP mode calculations at the current HEVC depth, and select the optimal prediction mode from these two modes. When it is predicted that the current CU does not need to be divided, follow the HEVC standard encoding process Optimal mode selection is performed, thereby reducing computational complexity in the mode selection process and improving the transcoding speed of the entire transcoding process.

The 2N×2N mode calculation refers to: indicating that the current coding unit can no longer be divided;

The SKIP mode calculation refers to: indicating that the current encoded macroblock does not need to encode prediction residual information;

The optimal prediction mode refers to: compare the rate-distortion of 2N×2N and SKIP modes, and select the encoding mode with the smallest rate-distortion cost;

The optimal mode selection refers to a process of comparing rate-distortion for all encoding modes in HEVC, so as to select the process with the least rate-distortion cost.

In summary, the advantages of this embodiment are:

1) Combined with the characteristics of the video itself, each sequence can be trained to obtain a model suitable for the current sequence.

2) According to the training model, predict the computational complexity of determining the optimal division of CU in the HEVC encoding process in transcoding, thus greatly reducing the computational complexity of transcoding while ensuring the prediction accuracy, saving transcoding time, and improving Transcoding efficiency.

3) Combined with the fast mode selection algorithm, the unnecessary calculation process is further reduced, and the transcoding speed is improved.

The algorithm proposed by the present invention is implemented on the basis of HEVC reference code HM10.1. The AVS-to-HEVC full-decoding full-coding transcoder is compared as a standard transcoder, and the effectiveness of the algorithm is evaluated by comparing three criteria, namely ΔT, ΔPSNR and ΔBR. It is calculated as follows:

ΔPSNR = PSNR _ref - PSNR _prop

ΔT, ΔPSNR and ΔBR represent the percentage of saved encoding time, PSNR reduction value and code rate growth rate, respectively.

The experimental results are shown in Table 1.

Table 1. Experimental results of AVS to HEVC transcoding algorithm based on support vector machine.

From the test results in Table 1, it can be seen that for the sequence of 1920×1080, the transcoding algorithm based on support vector machine can increase the transcoding speed by 63.68% on average, reduce the PSNR by 0.083dB on average, and increase the code rate by only 1.82 %, for the 1280×720 sequence, the transcoding speed is increased by 71.78% on average, the PSNR is reduced by 0.09dB on average, and the code rate is increased by 1.55%. Experimental results show that the algorithm of the present invention greatly improves the video transcoding speed while ensuring limited video quality degradation.

Claims (4)

1. AVS based on support vector machine to HEVC optimization video transcoding method, it is characterized in that, by collecting the feature vector of AVS code stream, and utilizing support vector machine to learn it and obtain training model, the AVS that will extract The eigenvectors are divided into two types: the CU at the corresponding position in HEVC or not. In the transcoding stage, the training model is used to predict whether the CU needs to be divided. When the current CU needs to be divided, 2N× is performed at the current HEVC depth. 2N mode and SKIP mode calculation, and select the optimal prediction mode from these two modes. When it is predicted that the current CU does not need to be divided, the optimal mode selection is performed according to the HEVC standard encoding process; The AVS feature vector is collected from the AVS code stream, including: macroblock coding mode, motion vector and transform coefficient; The optimal prediction mode refers to: compare the rate-distortion between 2N×2N and SKIP mode, and select the encoding mode with the smallest rate-distortion cost. 2. The method according to claim 1, wherein said macroblock encoding mode refers to: the mode information of each macroblock in the AVS; said motion vector refers to: the average motion of the macroblock in the AVS The size of the vector; the transform coefficient refers to: the number of non-zero coefficients in the discrete cosine transform coefficients of the AVS code. 3. The method according to claim 1 or 2, wherein the macroblock coding mode refers to: when the depth in HEVC corresponding to AVS is 0, each CU contains 16 macroblocks, so Contains 16 features; when the depth is 1, each CU contains 4 macroblocks, and correspondingly has 4 mode features; the motion vector refers to: the motion vector in AVS is in units of 8×8, A macroblock contains 4 motion vector basic units, and each motion vector is modulused separately, and then the average value of the four motion vector moduli in a macroblock is used as a feature. When the depth is 0, there are 16 motion features in the CU , when the depth is 1, it contains 4 motion features; the transformation coefficient refers to: DCT in AVS is also based on 8×8 blocks, and the number of non-zero coefficients in all DCT coefficients in a macroblock is used as a feature, When the depth is 0, it contains 16 DCT coefficient features, and when the depth is 1, it contains 4 DCT coefficient features. 4. The method according to claim 1, wherein the selection of the optimal mode refers to: performing a rate-distortion comparison on all encoding modes in HEVC, so as to select a process with the least rate-distortion cost.