Super-Resolution SAR Image Reconstruction via Deep Unfolding Network and Modified VGG Network

Synthetic aperture radar (SAR) as an advanced remote imaging system is affected by its inherent speckle noise, atmospheric disturbances and motion blurring factors. In this paper, the problem to improve the resolution of SAR is considered in a practical scenario, and we present a super-resolution (SR) reconstruction method based on deep unfolding network and a modified very deep convolutional (VGG) network. Specifically, the degradation process of the SAR image is first predefined to decompose it into a data subproblem and a prior subproblem. Then, two subproblems are solved by neural modules, resulting in an end-to-end trainable and iterative network. A modified VGG network is used for multiplicative to additive noise fitting as a trained model. During the experiments of real SAR images, our method performs clearer visual effects than other super-resolution reconstruction and achieves better performance on peak signal-to-noise ratio.