Super-resolution reconstruction of optical remote sensing images based on multiscale adaptive dynamic aggregation network
TL;DR
A Multi-scale Adaptive Dynamic Aggregation network (MADA-Net) is proposed, which effectively alleviates structural blurring and texture adhesion in subjective vision, and verifies its advancement and practical value in optical remote sensing image super-resolution tasks.
Abstract
The super-resolution reconstruction of optical remote sensing image is very important, which can break through the limitation of hardware physical resolution and improve the value of data application. To solve the problems of large scale span, complex compound degradation and difficulty in balancing global context and local details of single receptive field in remote sensing images, a Multi-scale Adaptive Dynamic Aggregation network (MADA-Net) is proposed in this paper. In which, a static multi-scale feature extraction module (SMFE) is constructed, which jointly extracts local details, regional contextual information and channel saliency features through a three-branch parallel structure. Then, Adaptive Dynamic Convolution (ADC) module is designed to dynamically adjust the convolution parameters and receptive field size according to the input content, so as to carry out the content adaptive modelling. The 4x super-resolution experiment is carried out on the public AID remote sensing dataset, the result shows that the average PSNR of MADA-Net reaches 29.82 dB and the SSIM reaches 0.7856, which is significantly better than the mainstream methods such as Bicubic, EDSR, RDN, and FMSR. It effectively alleviates structural blurring and texture adhesion in subjective vision, and verifies its advancement and practical value in optical remote sensing image super-resolution tasks.