Multi-scale residual reconstruction algorithm based on compressed sensing image

Existing compressed sensing image reconstruction algorithms based on deep learning have the problems of poor reconstructed image quality and blurred local areas. A multi-scale residual reconstruction network based on compressed sensing images is proposed. The proposed network consists of sampling sub-network and reconstruction sub-network.. First, the sampling sub-network performs compressed sensing sampling on the entire image through global convolution, and then obtains a preliminary reconstructed image through deconvolution; the reconstruction sub-network has multi-scale residual blocks with different sizes of convolution kernels, and each residual block has Different branches extract multiple image features. The experimental results show that the reconstructed image of the proposed network has better visual effects and clearer details. At the same time, the local information and texture structure of the image are enhanced, and the objective index of PSNR is significantly improved.