Multiwavelet Based Multibranch Skip Feature Fusion Network for Remote Sensing Image Super Resolution

In recent years, Convolutional Neural Network (CNN) based remote sensing image super-resolution (SR) has made great progress and attracted much attention. However, applying them to low computing capacity devices in the real world is challenging due to the vast number of parameters. This paper proposes a Multibranch Skip Feature Fusion Network (MSFFN) with minimal parameters and better restoration performance. The primary objective of the proposed work is to reconstruct high-resolution (HR) images in the multiwavelet domain. The proposed model utilizes the balanced Geronimo-Hardin-Massopus (GHM) Multiwavelet Transform (MWT) to decompose images into sixteen subbands, thereby enabling an efficient representation of fine edges and textures. A MSFFN is designed to aggregate hierarchical features across scales, while the Channel Splitting Distillation Block (CSDB) refines hierarchical features by separating and fusing contextual and high-frequency information. Furthermore, the Feature Enhancement Attention Module (FEAM) adaptively emphasizes salient features, ensuring precise detail recovery. Experiments conducted on three benchmark remote sensing image datasets demonstrate that the proposed MSFFN achieves superior reconstruction accuracy. The image of the Gangotri glacier captured on September 9, 2001, by NASA’s Terra satellite, is also included in the model’s performance analysis for real-time images with complex patterns and textures. The reconstructed images have better visual quality than the existing approaches with 719 k parameters.