Research On Video Coding Optimization For HEVC Based On Visual Saliency

High efficiency video coding is the development of H.264/AVC.It can further improve video coding efficiency with good video quality.Compared with H.264,intra prediction in HEVC increases angle directions and adopts a flexible coding structure,which improves the coding performance.However,these emerging technologies also increases the coding complexity.When watching the test sequence,people always pay attention to areas of high saliency,and ignore some corners and background information independently.In order to solve the problem of large computational complexity in HEVC encoder,the main content of this thesis is to build a saliency model by simulating the human visual system which extracts the visual information from test sequences and add the model into the intra prediction.The main work of this thesis includes the three following aspects:(1)Propose a saliency detection model combined global and local features: Firstly,the airspace global saliency map can be obtained by image signature.Secondly,we bulid luminance feature saliency map and texture feature saliency map by DCT low-frequency AC-DC coefficients from three channels of YUV video image.The airspace local salieny map is obtained by combining the luma features saliency map with the texture feature saliency map.Finally,the required saliency map is fused by the airspace global saliency map and the airspace local saliency map.(2)Propose an intra prediction optimization scheme based on saliency detection model.In the aspect of CU depth,we use the proposed saliency detection model to get the saliency values of all LCU blocks in the current frame,and apply the corresponding connection between saliency values and saliency level regions to obtain the depth range of the current LCU.In the aspect of the intra prediction mode selection,we get the main direction mode of the current PU based on Sobel operator and calculate all PU blocks saliency values in the current depth with the proposed saliency detection model.Then we choose a saliency region for current PU and get a suitable subset of candidate modes by the relationship between saliency regions and candidate mode subsets.At last we further filter the subsets which have both the primary direction and the subset of candidate modes for last RDO mode selection.(3)In the experiment,we use EWSSIM based on eye-tracking as the subjective quality evaluation standard and analyze our algorithm.Large numbers of experiments show that the proposed saliency model can accurately extract regions with high saliency in the video image.The intra-coding optimization scheme based on saliency can effectively reduce the coding time with good objective video quality and subjective video quality.