Dual-path denoising and feature-domain optimization network for image compressed sensing

Image Compressed Sensing (ICS) reconstruction based on deep unfolding network has achieved remarkable success. However, most of the ICS deep unfolding networks do not take full advantage of the critical information contained in the feature channels in image processing as well as the correlation between the local features of the image and the feature channels. To solve these problems, we designe an ICS deep unfolding network called dual-path denoising and feature-domain optimization network (DDFONet) using the idea of progressively optimizing image features. For feature extraction of images, we combine convolution and depth-separable convolution to capture local features of an image and deal with dependencies between image feature channels. We design a model based on the channel attention mechanism to focus on key features in image processing for optimal denoising of feature maps. The experimental results show that our proposed network gets good performance in image performance metrics such as SSIM and PSNR, as well as visual effects of several public benchmark datasets.