Learning a Deep ResNet for SAR Image Super-Resolution

To improve the performance of the traditional SAR image super resolution model, we propose a novel SAR super-resolution model based on the improved SRResnet. In the proposed framework, the batch normalization layer is removed. It can improve the details of image restoration. ReLU function is added to maintain the numerical properties. In addition, the pixel shuffler block is designed to upsampling the SAR images, which can reduce artifacts. We use a residual learning strategy to address the issue of vanishing gradient with the increasing of network depth. Compared with existing super resolution algorithms in the SAR images area, the proposed model achieves a good performance on both quantitative and visual assessments.