Gradient Prior Dilated Convolution Network for Remote Sensing Image Super-Resolution

Abstract

Super-resolution (SR) aims to recover a high-resolution image from a single or multiple low-resolution images, compensating for the limitations of satellite sensor imaging. Deep convolutional neural networks have made great achievement in remote sensing image SR. In this article, we propose a novel gradient prior dilated convolutional network (GPDCN) for remote sensing image SR, which obtains contextual spatial connections and alleviates structural distortions. The GPDCN comprises a multiscale feature extraction network and a feature reconstruction network. The former employs a double-path dilated residual block with dilation convolution to increase a receptive field, a global self-attention module to detect long-range reliance among image patches, and a gradient propagation network to extract high-level gradient information. The latter uses the mixed high-order attention module to reconstruct the feature by collecting the high-order characteristics of multiple frequency bands. Experiments with the Massachusetts_Roads and 3K VEHICLE_SR datasets demonstrate that the GPDCN outperforms recent techniques concerning both quantitative and qualitative measures.