Thermal radiance-inspired network for infrared small target detection

Infrared dim small target detection holds significant importance and has wide-ranging applications in many fields such as night search, rescue and environmental monitoring. However, infrared dim small targets are characterized by their small size, lack of distinct features, and low contrast, which pose challenges for existing deep learning methods, resulting in low detection accuracy and high computational costs. To address this problem, we propose a lightweight deep learning method named Thermal Radiance-Inspired Network (TRINet). Based on multi-directionality of the thermal radiance of targets, thermal radiance multi-directional (TRMD) module is designed to enhance target features during both encoding and decoding processes. Additionally, multi-scale characteristics of the thermal radiance are extracted by spatial feature modulation (SFM) module and cross-window attention (CWA) module with multi-scale features at the same resolution and different resolutions, respectively. Experimental results with comparison methods demonstrate that TRINet achieves state-of-the-art performance across various datasets. Meanwhile, TRINet achieves a detection speed of 25 fps on embedded system. The source codes will be available at https://xzbai.buaa.edu.cn/.