Research On Fast Algorithm Of AVS3 Based On Frequency Domain Coefficients

AVS(Audio Video coding Standard)is the second generation source coding standard with full independent intellectual property rights developed by China.In order to meet the growing demand of China’s video industry for video coding efficiency,three generations of video coding standards have been developed and formulated.Now the latest thirdgeneration AVS3 video coding standard has completed two stages of development tasks,and has significantly improved the coding efficiency compared to previous standards,but due to the introduction of a large number of new encoding tools,its coding complexity has also increased exponentially.In this regard,it is significant to conduct research on fast algorithms for AVS3 video coding.In this thesis,we investigate the Coding Unit(CU)classification and intra-frame predictive coding of AVS3 from the perspective of frequency domain,and our main work is as follows.To address the problem that the coding complexity increases dramatically due to the multiple CU division methods added in AVS3,this thesis starts from the physical meaning of the Discrete Cosine Transform(DCT)coefficients in the frequency domain,finds that the DCT coefficients of image blocks are closely related to the image texture,and verifies experimentally that the texture complexity of image blocks is related to the low frequency coefficients,and the horizontal or vertical textures are related to the first column or first row of DCT coefficients.Based on this,this thesis proposes a fast algorithm based on DCT coefficients,which calculates the complexity and horizontal and vertical texture intensity of image blocks,and guides the CU division accordingly,terminates the CU with low complexity in advance to continue by setting two thresholds,and skips the vertical(horizontal)division of the CU with high horizontal(vertical)texture intensity,which can reduce the cost of computing CU encoding the number of traversals needed,and thus speed up the encoding speed.Finally,the performance of the algorithm is tested and the optimal threshold value is found.The test results show that the algorithm optimally saves an average of 30.6%of encoding time in the AI(ALL INTRA)mode of the reference software platform HPM 14.0,while the average encoding loss of the luminance component is only 0.58%.In addition,to address the problem of increasing complexity of intra prediction coding in AVS3,this thesis starts from the original Rough Mode Decision(RMD)algorithm of AVS3 and finds that the orientation of the best prediction mode can be roughly located by using the texture angle instead of the traversal of the original 4-fold angle mode.In this thesis,we investigate the texture angle of the image block in two dimensions:DCT coefficients and spatial domain gradients.By designing a texture angle calculation tool to assist the exploration,we find that the relative intensities of horizontal and vertical textures calculated by DCT coefficients or Sobel operator can reflect the approximate texture angle of the image.Based on this,this thesis proposes a fast algorithm for intra prediction pattern selection by mapping the obtained texture angles to six intervals of angle prediction patterns,replacing the initial candidate pattern set of 4 times the angle in the first step of RMD,and reducing the coding complexity by reducing the number of patterns that need to be traversed and calculated for the subsequent process accordingly.Finally,the performance of the two algorithms is tested and compared.The test results show that in the AI mode of the reference software platform HPM 14.0,the frequency domain algorithm has an average coding time saving of 29.5%and an average Ycomponent loss of 0.64%,and the spatial domain algorithm has an average coding time saving of 31%and an average Y-component loss of 0.66%.