| As a typical 3D video representation,multiview video records scene data from multiple viewpoints and can provide users with interactive experiences such as viewpoint selection and scene roaming.Therefore,multiview video is widely used in various new video applications,such as virtual reality,augmented reality and six degrees of freedom system.With the introduction of multiple viewpoints,the data volume of multiview videos has also increased significantly,which brings huge challenges to the storage capacity of storage devices and the transmission capacity of transmission media.In order to efficiently compress multiview video,the international video coding standardization organization has formulated a multiview video coding standard and designed in-loop filtering technology to reduce coding distortion and improve compression efficiency.In order to further improve the coding efficiency of multiview video,this thesis has carried out research on deep in-loop filtering technology for multiview video coding,and constructed the quality enhancement method of reconstructed frames based on the spatial,temporal,and inter-view correlation of multiview video,which effectively improves the compression efficiency of multiview video coding.This thesis proposes a deep in-loop filtering method based on inter-view information fusion.According to the inter-view correlation of multiview video,a quality enhancement network based on the inter-view information fusion is constructed to improve the quality of reconstructed videos.First,based on the inter-view correlation,reconstructed frames from adjacent viewpoints are utilized as the inter-view reference frames to provide effective reference information.Then,in order to sufficiently enhance the reconstructed frame,a multi-level receptive field fusion module is designed to effectively extract multi-scale features and fuse inter-view information.Finally,as a deep learning-based filter,the proposed method is integrated into the 3D-HEVC coding platform.Experimental results show that the proposed method can effectively improve the coding efficiency of multiview video.This thesis also proposes a deep in-loop filtering method based on multi-domain correlation learning and partition constraint.According to the spatial,temporal,and inter-view correlations of multiview video,a quality enhancement network based on multi-domain correlation learning and partition constraint is constructed to enhance the quality of the reconstructed video more sufficiently.First,in order to effectively restore the high-frequency details lost in the distorted image,a multi-domain correlation learning module is designed to mine temporal reference information and inter-view reference information.Then,a partition-constrained reconstruction module is proposed.This module uses the block partition characteristics of video coding to further reduce coding distortion,by designing a partition loss function.Finally,the proposed method is integrated into the 3D-HEVC coding platform,and the optimal mode selection is realized according to the rate-distortion optimization criterion.Experimental results show that the proposed method can significantly improve the coding efficiency of multiview video. |
