An efficient HEVC intra frame coding based on convolutional neural network

• Main Authors: Yueh-Ju Lu, 呂岳儒
• Other Authors: Chou-Chen Wang
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/u373by

碩士 === 義守大學 === 電子工程學系 === 107 === Convolutional neural network (CNN) has been developed rapidly in deep learning areas, and has become a hot studying topic in image applications. High efficiency video coding (HEVC) is the newest video coding standard. The HEVC standard can achieve high coding efficiency with a lower bitrate for intra frame coding. However, it still needs many bits to finish best rate-distortion (R-D) curve. Since there are only 35 directions prediction modes provided in intra prediction module (IPM), HEVC occurs a large distortion when the image contents are out of these prediction directions. In other words, HEVC can achieve high coding performance when the image contents match these 35 prediction directions. In order to obtain a better R-D curve, Zhang et al. [3] recently proposed a simple CNN (S-CNN) to improve the encoding performance of HEVC. The S-CNN consists of super-resolution CNN (SRCNN) [7] and ResNet [8] with two layer networks. The S-CNN can precisely predict the residual information of coding tree unit (CTU) and achieve a better R-D performance for HEVC encoder. However, S-CNN has to consume more time to encode intra frame coding since it needs to perform more CNN enhancement mode. In order to further speed up S-CNN based intra frame coding, we propose an early termination algorithm to skip CNN. Because the natural images have high spatial correlation, we find the mean square errors (MSE) of reconstructed CTU exist high spatial correlation at HEVC encoder. Therefore, a dynamic threshold of MSE is set according to four neighboring encoded CTU blocks to evaluate whether the current reconstructed CTU is useful for the CNN enhancement mode. Simulation results show that both our proposed method and S-CNN can reach better R-D curves. On the other hand, although our proposed method increases 0.9% and loses 0.05 dB in average BD-BitRate and BD-PSNR as compared with S-CNN, respectively. However, we can achieve faster HEVC encoding process than S-CNN by reducing time increase ratio (TIR) about 13% on an average.