Research On Deep Network-Based Compressed Sensing For Image Compression Using Algorithm Unfolding

In the past few years,digital information has been growing exponentially due to the rapid development of technologies such as artificial intelligence,cloud computing,and big data.The traditional Nyquist-Shannon sampling theorem requires signals to be sampled at a rate higher than twice the signal bandwidth for accurate reconstruction.However,sampling and storing large-scale data at high sampling rates has become increasingly impractical.Compressed sensing technology has emerged as a solution to this problem by reducing the sampling amount and achieving signal compression and recovery.Many practical signals have sparsity or approximate sparsity,which can be exploited by compressed sensing algorithms to reconstruct signals.Deep learning-based compressed sensing has become a hot research topic in recent years due to its superior reconstruction quality and generalization ability.However,existing deep learning models suffer from block effects,high sampling rates,lack of interpretability,and insufficient flexibility.This paper aims to address these problems by proposing algorithms for reconstructing high-quality signals from severely missing measurement data.The main contributions of this paper include:(1)A deep unfolding model based on the quadratic optimization perspective of the Split Augmented Lagrangian Shrinkage Algorithm(SALSA),named SALSA-Net,is proposed.The model is divided into three parts: sampling,initial reconstruction,and deep reconstruction.The sampling part is based on convolutional sampling,which uses the sparsity and texture features of the signal to improve the reconstruction quality without manual intervention.The reconstruction part maps the updates of the SALSA algorithm onto a deep network and decomposes the model into solving two subproblems and an auxiliary update module,inheriting the good interpretability of the SALSA algorithm.All the parameters of the model are optimized based on end-to-end learning,and the parameters are constrained forward to ensure correct convergence and improve network stability.Experimental results show that compared with existing deep unfolding methods,the proposed method can effectively reconstruct image detail texture information and has good reconstruction visual effects.(2)A deep unfolding model named SALSA-Net+ is proposed,which combines the nonlocal self-similarity prior and the SALSA algorithm.During the sampling process,reference sampling values are obtained by utilizing the correlation between different blocks in the image,and the block effect of block-by-block reconstruction is eliminated by using the correlation and texture information between different regions,resulting in a better initial reconstructed image.In the deep reconstruction part,a dynamic adaptation strategy is introduced,which uses a few number of parameters to realize sampling and reconstruction of images at different sampling rates for a single model.Experimental results show that SALSA-Net+ can achieve high-quality and flexible sampling and reconstruction,and has more prominent reconstruction performance at low sampling rates.