Remote Sensing Image Super-Resolution Using Feature Grouped Multi-Scale Network

Abstract. With the rapid development of deep learning technology, remote sensing image super-resolution has made remarkable progress. Remote sensing images usually contain multiple objects with different scales, making it crucial to adopt multi-scale feature extraction methods. However, the existing multi-scale modules introduce a large number of parameters due to performing convolution operations with different kernel sizes on the input feature maps separately. To address this issue, this paper proposes a Grouped Multi-scale Feature Extraction (GMFE) module. By applying grouped convolutions along the depth dimension of the input feature maps, the number of parameters is effectively reduced. On this basis, we design a Feature-grouped Multi-scale Super-Resolution (FMSR) network. We propose an Edge Enhancement (EE) module integrated into the network to sharpen edges and enhance the visual quality of the image. Additionally, we introduce an arbitrary scale upsampling module, enabling a single trained model to perform image super-resolution reconstruction at arbitrary scales. Extensive experiments on the UC Merced and RSSCN7 datasets demonstrate that the proposed FMSR network achieves superior performance in both quantitative metrics and visual quality.