Compressive Sensed Video Recovery Using Iterative Thresholding with Quality Enhancing Neural Networks

This paper is dedicated to application of video enhancing neural networks for recovery perfomance improvement of traditional (neural networks free) iterative shrink-age/thresholding algorithms (ISTA). First, we show that such a network can be periodically used at some iterations of ISTA. Herewith, the quality imrovement can be achieved especially for video sequences having a high motion activity compressed at low bit rates (or sensing rates). Second, we show how this approach can be embedded into CS-JPEG codec which is based on the compressive sensing framework. Finally, experimental results obtained at low bit rates show that comparing to original CS-JPEG codec the proposed approach can improve Peak Signal-to-Noise Ratio (PSNR) up to 2 dB, and Video Multimethod Assessment Fusion (VMAF) up to 17 scores. Comparison with JPEG and H.264/AVC Intra is also provided.