Dictionary learning on l1-norm fidelity for non-key frames in distributed compressed video sensing

Abstract

Distributed Compressed Video Sensing (DCVS), which consists of Compressed Sensing (CS) and Distributed Video Coding (DVC), is an encoding scheme transferring computational burden from encoder to decoder. By assuming that given signals are sparse, the CS enables accurate decoding only referring low dimensional observations which are obtained by low-rank random projection of original signals. The DVC divides image sequences into key and non-key frames and regards the decoding of the non-key frames as error correction using the key frames. The quality of the non-key frames depends on the design of the dictionaries. Then, many studies optimize dictionaries with convex optimization solvers for functions consisting of the weighted sum of two terms: l 2 -norm error estimation term and l 1 -norm regularization term. This paper proposes to use l 1 -norm error instead of l 2 -norm to increase the robustness against outliers. We apply ADMM (Alternating Direction Method of Multipliers), which is a convex optimization solver, to minimization the cost function. Simulation results show the proposed method generates better Quality images than the conventional method.