EPSFusion Network for Infrared and Visible Image Fusion-Based Semantic Segmentation

Infrared and visible image fusion has attracted considerable interest in computer vision, particularly due to its valuable applications in object detection and autonomous driving tasks. In this study, we propose the EPSFusion network, utilizing a ResNet101 backbone, to enhance semantic segmentation performance for fused infrared and visible images. To comprehensively improve local and global feature representation, our designed network integrates a dual-attention module (DAM) focusing on channel and spatial dimensions, along with a Transformer encoder. The DAM module leverages channel and spatial attention mechanisms to refine local feature extraction, while the Transformer encoder effectively captures long-range dependencies via self-attention, thus optimizing intermodal information fusion across modalities. The network's effectiveness is validated through training and testing on the MSRS dataset, with ablation experiments conducted to verify each component's contribution. Experimental outcomes reveal that our proposed approach achieves a mean Intersection over Union (mIoU) of 73.08%, surpassing the baseline model by 1.27%.