Super-Resolution of Remote Sensing Images based on a Deep Plug-and-Play Framework
Hongyuan Tao · 2020
Single image super-resolution (SISR) based on deep neural network (DNN) has been widely studied in recent years as a crucial technique for remote sensing (RS) applications. However, owing to the complexity and diversity of ground objects, there remains fundamental challenges to reconstruct a high-resolution (HS) RS image from a low-resolution (LR) RS image, especially with blur. In this paper, I propose a deep plug-and-play residual network, namely DPSRResNet, which can reconstruct high-quality HR RS images from LR SR images with Gaussian blur kernels via a deep plug-and-play framework. Specifically, a degradation model from the DPSR framework is given to utilize matured deblurring methods. Moreover, I adopt a deep plug-and-play algorithm to optimize the energy function, which allows plugging any super-resolver with a prior term. The proposed DPSRResNet is used as the crucial super-resolver for the framework, and a series of experimental results are presented to demonstrate the effectiveness of the proposed method on RS images.