PLV-CSNet: Projected Landweber Variant unfolding network for image compressive sensing reconstruction

Due to the powerful learning capability and fast processing speed of deep neural networks, a series of pure data- driven and deep unfolding networks for image reconstruction have emerged, achieving improved reconstruction quality. These reconstruction networks typically employ convolutional neural networks or residual neural networks to extract high-dimensional features of the dominant structure component. However, the edge and texture components in multi-dimensional features as well as the measurement residual generated at each iteration during the unfolding procedure are often neglected, which would affect the quality of image reconstruction. In this paper, a projected Landweber variant unfolding network (PLV-CSNet) is proposed for image compressive sensing reconstruction. A PLV algorithm is investigated and then unfolded to the PLVBlock, which consists of a thresholding module (TSM) and a progressive projecting module (PPM). The TSM utilizes the dense block to fuse multi-dimensional image features and the soft thresholding to eliminate image noise. The PPM combines the approximate message passing algorithm with deep neural networks to compute the projections of the approximation solution for images, as well as calculate the measurement residual generated during each iteration. Furthermore, a residual integration module (RIM) is designed to employ the measurement residuals to reconstruct the image residual which are then flexibly supplemented back into the reconstructed image. The effectiveness of PLV-CSNet is demonstrated in four standard benchmark datasets, and comparisons with classical image compressive sensing reconstruction networks show that our network could achieve higher reconstruction accuracy. Codes are available at: https://github.com/junp-hao/PLV-CSNet.