MBMS-Net: A Multi-Branch and Multi-Scale Perception Infrared Small Target Detection Network

Infrared small target detection (IRSTD) has seen broad adoption in military and civilian applications, where it plays an important role. The small target size and complex background often result in feature degradation and class imbalance during detection. To overcome the previously discussed limitations, this study proposes a novel method named Multi-Branch and Multi-Scale Perception Infrared Small Target Detection Network (MBMS-Net). Specifically, a Multi-Branch Segmentation Fusion Module (MBS) is introduced to aggregate features across multiple resolutions, improving network performance in recognizing and retaining small targets, while supporting hierarchical feature extraction and cross-branch data exchange. Subsequently, a Multi-Scale Dilated Fusion Module (MSDF) enables contextual perception under varying receptive fields and facilitates feature fusion across scales. Furthermore, a Hybrid Attention Module (HAM) is applied to dynamically focus on target regions while suppressing background interference. The performance of the proposed network has been thoroughly evaluated on the SIRST-AUG dataset, showing notable improvements in detecting infrared small targets and achieving superior results compared to existing deep learning and traditional approaches.