Compressive sensing with adaptive pixel domain reconstruction for block-based video coding

This paper presents a new look at image/video compression from the compressive sensing´s perspective. Quantization in video compression can be regarded as a subsampling process where the signal is mapped into predefined levels. We view the problem of signal reconstruction from its quantized signal vector as a compressive sensing recovery problem where the quantized coefficients are subsampled measurements. Based on this observation, we propose a novel method of image/video coding that employs an adaptive Total-Variation (TV) minimization in the pixel domain to recover the gradient-sparse image blocks from their quantized transform coefficients. We further increase the coding efficiency by encoding only a subset of the transform coefficients and discard the remaining ones. Experiment results show that the proposed framework is efficient with gradient-sparse video signals and outperform the video compression standard H.264/AVC by up to 7% of bitrate reduction.