A Parallel and Optimized Image Compressed Sensing Solution

Abstract

Although there have been significant processor technology enhancements in terms of speed, data compression algorithms still do not accomplish the required task in a convenient time for voluminous data. The parallelism of the compression process could significantly improve not only the processing time but also the quality of the solution. In this paper, we propose a new Compressed Sensing (CS) solution based on Parallel CSNet (PCSNet) while integrating the PSNR loss to improve the image reconstruction process. The proposal results in distributing data to be processed to different cores of the machine and could benefit a multitude of applications including autonomous driving, medical imaging, and face detection. At its core, PCSNet uses a parallel convolutional neural network that includes a sampling subnetwork and a reconstruction subnetwork. These two networks learn the sampling matrix from the input image then the reconstructed image from the CS measurements. Thus, the training is done in a parallel way in each iteration then optimized at the master level using PSNR loss in an end-to-end learning process. Obtained experimental results outperform state-of-the-art approaches in terms of both image reconstruction quality and processing time.