Adaptive Block Partitioning And Inter Prediction

In the information age,digital multimedia plays an increasingly important role in people's daily lives.Digital video is the most important multimedia carrier and serves as the critical way for humans to acquire information and perceive the world.Due to the pursuit of visual quality and media interaction,the video resolution and framerate increase consistently.Moreover,new video applications have been developed,including virtual reality,point cloud,holographic projection and so on.As the huge increment of video data and new media applications,higher requirements and new challenges have been put forward for the vidoe coding.However,current video coding technology cannot perform well for these changes,and the relative rigescent mechanisms limit the further improvement of compression efficiency.In the thesis,we take the adaptive decision as the starting point,focusing on the partitioning and prediction in video coding.The major contributions of the thesis lie in the following aspects:· Local-constrained based adaptive partitioning structure.The current standard adopts multi-types partitioning tree structure,including the quadtree partitionnig,triple tree partitioning and binary tree partitioning.However,the multi-types partitioning tree is not flexible enough for some image areas,while the partitioning modes overflow for other areas.To address this drawback,local-constrained based adaptive partitioning structure method is proposed.On one hand,the partitioning parameters can be derived on the decoder-site by utilizing the information of spatial-temporal reconstructed blocks.The partitioning structure and depth are adaptively adjusted based on the video content characteristics,especially for the smooth areas.On the other hand,according to the partitioning redundancies among different partitioning types,the impact of partitioning on the encoding order is analysed.Based on that,local constraint is applied on the partitioning types and the redundancies are reduced.· Confidence-interval based partitioning decision.Multi-types partitioning tree structure requires multi-iterations on each depth,bringing huge computational complexity.To fast the partitioning decision,we first develop the rate-distortion model based on the motion divergence field.Based on this model,the rate-distortion cost can be fast estimated.Then,the relationship between the optimal partitioning mode and the esitimated rate-distortion cost is mapped into probability interval.Furthermore,the confidenceinterval based partitioning decision algorithm is proposed.The partitioning mode in the confidence interval is classified as a reliable mode,and the partitioning mode outside the confidence interval is classified as an unreliable mode.By this way,the proposed method reduces the encoding complexity significantly while revserves the effective partitioning.· Content-oriented adaptive partitioning of the prediction units.The flexibility of multi-types partitioning tree structure has been improved,but still cannot obtain satisfied prediciton around the boundaries of moving objects.In this view,reference segmentation based adaptive partitioning of the prediction units algorithm is proposed.First,a motion vector is signalled to point out the reference block,which provides the boundary information.Then,the reference block is partitioned into three zones and each zone adopts different prediction methods.The foreground zone is motion compensated by the implicitly multi-hypothesis prediction.The background zone is motion compensated based on the local motion field.Regarding the edge zone,it is compensated by the weighted prediction.The proposed algorithm realizes object-oriented adaptive partitioning and meanwhile avoids huge overheads.Simulations show that the algorithm can improve the subjective and objective performance.· Adaptive motion vector resolution prediction based on rate-distortion optimization.Partitioning aims to obtain better prediction.Moreover,prediciton performance is also influenced by the description accuracy of motion information.As the motion vector resolution increased,the motion can be captured more precisely.However,the coding rate of motion vector also increases.To achieve better trade-off between the prediction accuracy and the coding rate,we develop the rate-distortion models in terms of motion vector resolution.The models reveal that the motion vector resolution prediction is influenced by the texture and motion in video content,and meanwhile influenced by the coding environment.Based on that,the influencing factors of the motion vector resolution selection are selected and trained into the decision tree to predict the optimal motion vector resolution for each frame.This method can achieve better trade-off between the prediction accuracy and the motion vecotr coding rate,and hence improve the compression efficiency.As summary,the thesis focus on the partitioning and prediction in video coding,and investigate the influence of video content characteristics in the view of rate-distortion optimization.Adaptive partitioning structure,confidence-interval based partitioning decision,adaptive partitioning of the prediction units and adaptive motion vector resolution prediction algorithms are proposed.This thesis explore the potential of adaptive mechanisms in video coding.The proposed algorithms can greatly improve the coding efficiency and decrease the computational complexities,revealing new ideas to research and application.