A dual-path feature reuse multi-scale network for remote sensing image super-resolution

Deep neural networks have achieved significant success in the super-resolution of remote sensing images. However, existing deep learning models still suffer from the issue of blurry pseudo-artifacts when restoring high-frequency details and textures. In this paper, a novel dual-path feature reuse multi-scale network (DFMNet) is proposed to more effectively utilize multi-scale features in remote sensing images, enhancing the detailed information in the restored images. Specifically, the designed dual-path feature reuse module adopts a symmetrical dual-path structure, with each path composed of convolutional layers of different sizes. This module enables deep feature reuse and multi-scale aggregation, improving the network's ability to handle and restore high-frequency details in the images. Furthermore, a cross-attention module is introduced to facilitate deep interactive fusion of multi-scale image features produced by the encoder output. Comparative experiments conducted on challenging UCMerced and AID remote sensing datasets demonstrate that the proposed DFMNet achieves superior performance in both objective and subjective evaluations.