Cascade Scale-Aware Distillation Network for Lightweight Remote Sensing Image Super-Resolution

Recently, convolution neural network based methods have dominated the remote sensing image super-resolution (RSISR). However, most of them own complex network structures and a large number of network parameters, which is not friendly to computational resources limited scenarios. Besides, scale variations of objects in the remote sensing image are still challenging for most methods to generate high-quality super-resolution results. To this end, we propose a scale-aware group convolution (SGC) for RSISR. Specifically, each SGC firstly uses group convolutions with different dilation rates for extracting multi-scale features. Then, a scale-aware feature guidance approach and enhancement approach are leveraged to enhance the representation ability of different scale features. Based on SGC, a cascaded scale-aware distillation network (CSDN) is designed, which is composed of multiple SGC based cascade scale-aware distillation blocks (CSDBs). The output of each CSDB will be fused via the backward feature fusion module for final image super-resolution reconstruction. Extensive experiments are performed on the commonly-used UC Merced dataset. Quantitative and qualitative experiment results demonstrate the effectiveness of our method.