Note (General)type:Thesis
A series of video codecs, combining encoder and decoder, have been developed to improve the human experience of video-on-demand: higher quality videos at lower bitrates. Despite being at the leading of the compression race, the High Efficiency Video Coding (HEVC or H.265), the latest Versatile Video Coding (VVC) standard, and compressive sensing (CS) are still suffering from lossy compression. Lossy compression algorithms approximate input signals by smaller file size but degrade reconstructed data, leaving space for further improvement. This work aims to develop hybrid codecs taking advantage of both state-of-the-art video coding technologies and deep learning techniques: traditional non-learning components will either be replaced or combined with various deep learning models. Note that related studies have not made the most of coding information, this work studies and utilizes more potential resources in both encoder and decoder for further improving different codecs.In the encoder, motion compensated prediction (MCP) is one of the key components that bring high compression ratios to video codecs. For enhancing the MCP performance, modern video codecs offer interpolation filters for fractional motions. However, these handcrafted fractional interpolation filters are designed on ideal signals, which limit the codecs in dealing with real-world video data. This proposal introduces a deep learning approach for all Luma and Chroma fractional pixels, aiming for more accurate motion compensation and coding efficiency.One extraordinary feature of CS compared to other codecs is that CS can recover multiple images at the decoder by applying various algorithms on the one and only coded data. Note that the related works have not made use of this property, this work enables a deep learning-based compressive sensing image enhancement framework using multiple reconstructed signals. Learning to enhance from multiple reconstructed images delivers a valuable mechanism for training deep neural networks while requiring no additional transmitted data.In the encoder and decoder of modern video coding standards, in-loop filters (ILF) dedicate the most important role in producing the final reconstructed image quality and compression rate. This work introduces a deep learning approach for improving the handcrafted ILF for modern video coding standards. We first utilize various coding resources and present novel deep learning-based ILF. Related works perform the rate-distortion-based ILF mode selection at the coding-tree-unit (CTU) level to further enhance the deep learning-based ILF, and the corresponding bits are encoded and transmitted to the decoder. In this work, we move towards a deeper approach: a reinforcement-learning based autonomous ILF mode selection scheme is presented, enabling the ability to adapt to different coding unit (CU) levels. Using this approach, we require no additional bits while ensuring the best image quality at local levels beyond the CTU level.While this research mainly targets improving the recent video coding standard VVC and the sparse-based CS, it is also flexibly designed to adapt the previous and future video coding standards with minor modifications.
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2023-03-03T17:22:19+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション