Infrared dim and small target detection model based on improved edge-aware gating and three-path adaptive fusion

As a vital research direction of computer vision in the fields of national defense and civil monitoring, infrared dim and small target detection acts as a key technical support for tasks such as long-range early warning, maritime surveillance and environmental perception of unmanned equipment, which is of great significance for improving the environmental adaptability and task response accuracy of perception systems. Aiming at the problems of low signal-to-noise ratio, background clutter interference and inconspicuous target features in infrared dim and small target detection under complex scenarios, an improved infrared dim and small target detection model with edge-aware gating and three-path adaptive fusion was proposed. Firstly, a U-shaped encoder-decoder architecture was adopted to realize multi-scale semantic fusion, with Swin-Transformer as the backbone to uniformly model local details and long-range dependencies. Then, to effectively suppress background noise and enhance target edge responses, a collaborative module integrating the recursive enhanced edge-semantic gated contextual layer (REES-GCL) and atrous spatial pyramid pooling (ASPP) was improved and designed, which ccould adaptively suppress complex background clutter and accurately extract target edge information. Finally, addressing the issues of significant semantic gap and lack of targeted weight allocation in traditional feature fusion, a gated dual-stage attention fusion (GDSAF) module was improved and designed to achieve the precise collaboration of three-path features. Experimental results show that the proposed model achieves a significant improvement in robustness under complex backgrounds and low-contrast scenarios, with the detection precision and recall rate increased by 7.3% and 3.1% respectively, compared with the state-of-the-art comparison model, which verifies the effectiveness of the collaborative mechanism of REES-GCL and ASPP as well as the dual-stage attention fusion module.