Research On Compressed Sensing Image Reconstruction Algorithm Based On Neural Network

With the increasing demand for image quality in various applications,image processing involves an increasing amount of data.The sampling compression transmission method guided by the Nyquist sampling theorem brings more and more pressure to the sampling and storage devices in image processing.Compressed Sensing(CS)performs uncorrelated measurements on the signal at a sampling rate much smaller than the Nyquist sampling theorem,and reconstructs the original signal with high probability by solving the nonlinear optimization problem.From traditional Hand-Designed(HD)methods to data-driven Neural Network(NN)based methods,people have done extensive research on CS image reconstruction,and are looking for fast and efficient CS image reconstruction algorithms.HD has theoretical convergence guarantees,but there are many manually selected parameters.NN can adaptively select parameters and learn real signal priors from a large amount of training data,but based on the operation of black box,there is no theoretical convergence guarantee.It was recently demonstrated that iterative sparse-signal-recovery algorithms could be unrolled to form interpretable deep networks.This mixing hand-designed and data-driven method have great research potential and are expected to further improve the performance and efficiency of the CS image reconstruction algorithm.These methods first use expert knowledge to set up a recovery algorithm and then use training data to learn priors within this algorithm.Such methods benefit from the ability to learn more realistic signal priors from the training data,while still maintaining the interpretability and guarantees that made hand-designed methods so appealing.Following are the specific research contents:(1)This paper studies the denoising-based CS image reconstruction.Since direct end-to-end training of deep networks is difficult,Approximate Message Passing(AMP)can naturally combine denoising,decoupling deep networks layer by layer,and transforming complex reconstruction problems into different level denoising problems.This research inspired the design denoising-based reconstruction algorithms of high performance.(2)Designed a feature-aware adaptively denoiser-selection strategy for CS image reconstruction.The Approximate Message Passing(AMP)network provides a natural way to combine a denoiser.The performance of the denoiser directly determines the quality of image reconstruction.However,there is currently no denoising algorithm that can be universally applied to all types of images.Therefore,the algorithm that with fixed denoiser cannot adapted to all types of images.(3)A compressed sensing image reconstruction algorithm combining multiple de-noises and multiple priors is proposed.More priors are integrated into the CS imagereconstruction algorithm,which effectively improves the flexibility of the algorithm.It is found that the reconstruction path modification and reconstruction mapping will greatly affect the quality of deep AMP network reconstruction.Therefore,the introduction of NN to maximize the use of image local prior to correct the reconstruction path and introduce the image non-local self-similarity Validation is used to reconstruct the map.Compared with the existing high performance reconstruction algorithm,the algorithm can effectively improve the performance gain of 1-2 d B under the condition of 0.3-0.5 compression ratio.