Dictionary Learning Based Two-stage Near-lossless Video Compression

Traditional hybrid video coding framework using block based predictive coding and transform coding, such as the High Efficiency Video Coding (HEVC), cannot further dig out the redundancy remained in quantized transformed residual, causing extra bits consumption. Measured by rate-distortion (R-D) performance, the problem of higher bits consuming can be solved reversely by video quality enhancing. In this work, we proposed a video coding scheme that solve the problem by enhancing the reconstructed video quality using supplementary information from further compressed quantization error. Aiming at better R-D performance for near-lossless video coding, we propose a novel video coding scheme using a two-stage framework that extracts quantization error as complementary information which is compressed using dictionary learning and sparse representation. The employed over-complete dictionary is learned through K-SVD with orthogonal matching pursuit (OMP) for sparse representation. Statistically reduandancy is further removed by a modified context-adaptive binary arithmetic coding (CABAC) with adaptive context models. This approach not only retains the advantages of the traditionally encoder for lossy compression but also exploits the redundancy in the quantization error to achieve high-quality near-lossless compression. Experimental results demonstrate that our method significantly outperforms traditional HEVC lossy encoder with over -20% BD-BR on average at high bitrate range for near-lossless coding, while the method is also proved to be efficient at low bitrate range achieving over -50% BD-BR on average with average PSNR over 41dB, which retains near-lossless performance.