Aadptive Block Ordering And Non-Rectangular Block Prediction For Video Intra Coding

With the development of communication technology and electronic equipment,various video applications have already entered into every sphere of people's life and work.At the same time,people's demand for high-quality video with ultra-high defini-tion,high frame rate and high color depth,is gradually increasing.However,the videos of high quality have a larger amount of data meaning greater need for video compression technology.Therefore,high efficient video coding standard(High Efficiency Video Coding,H.265/HEVC)was released in 2013.The next generation standard(Versital Video Coding,VVC)is also under way.Video compression is achieved by removing redundancy data.Intra predictive cod-ing,in which prediction values are generated by reconstructed pixels in current image,mainly eliminates the spatial redundancy.Then the predicted residual rather than the original image will be transformed,quantified and encoded into bitstream.Therefore,raising the accuracy of intra/inter prediction can improve the coding efficiency.On the one hand,the existing video coding standards adopt fixed coding order,raster and Z-scan order.Only the top and/or left neighboring reconstructed pixels can be used,which only incurs accurate prediction for blocks with the direction of coding order,topleft-to-bottomright.Therefore,the fixed coding order cannot deal with the rich texture.On the other hand,the images can be considered as piecewise smoothing,which consists of regions with different characteristics.Besides,coding blocks are divided into rectan-gular subblocks in tree-like structure.Blocks within the same characteristic region may encode the same information,and blocks containing boundary regions only use one prediction model,causing ineffienct coding and incorrect prediction.Obviously,this rectangular prediction blocks can not deal with the irregular boundary area effectively.In this thesis,two aspects are studied to improve the accuracy of intra prediction:first,the coding order of blocks is changed adaptively to deal with various texture,that is,local spatial correlations;Second,a non-rectangular block prediction method based on convolution neural network is proposed to improve the coding efficiency of boundary regions.Specifically,this thesis has the following innovations:1.Based on quadtree of HEVC and multi-tree of VVC partition structure,the adap-tive block coding order is proposed.The optimal coding order is selected accord-ing to the image content so that different local spatial correlations can be dealt with the reference pixels at different positions;According to the hierarchical and flexible coding order,this thesis proposes an improved prediction method,which expands the range of angle modes without adding new prediction model;Also,this thesis studies the relationship between coding performance and complexity to find a better compromise.Experimental results shows that the method is able to improve the intra prediction accuracy and achieves on average 2.6%-3.1%BD-rate reduction.2.A non-rectangular block prediction method based on convolution neural network is proposed.According to the geometric structure of the coding blocks,the pre-diction blocks are divided into two non-rectangular parts.Different parts use different prediction models and obtain the overall prediction finally.In view of the shortcomings of the above method,an improved training method is proposed:predicting the pixel value directly instead of predicting the partition method first,which further improves the coding efficiency.Experimental results shows that the method is able to improve the intra prediction accuracy and achieves on av-erage 0.6%-1.2%BD-rate reduction...