Refining False Detections: An Efficient Network Focused on Error-Prone Regions in Infrared Small Target Detection

Deep learning-based methods have achieved remarkable performance in infrared small target detection (IRSTD). However, existing approaches have not fully captured the intrinsic characteristics of this task. Experimental analysis shows that infrared small targets have relatively simple features, and the main challenge lies not in localization, but in accurately identifying the boundary pixels of patch-like and point-like targets, which are often misclassified or overlooked by current methods. Based on these observations, this paper proposes a two-stage IRSTD network, RefineFD, which detects targets from coarse segmentation to fine refinement. Unlike conventional two-stage methods based on candidate regions (e.g., RPN), the proposed approach directly focuses on uncertain regions prone to false detections, and its second stage does not rely on the outputs of the first stage. This design ensures a fixed and lower computational cost while effectively correcting false detections from the first stage. Specifically, the first stage uses a shallow and narrow network to achieve coarse segmentation by exploiting the simplicity of small target features, while the second stage further refines the results by leveraging the strong correlation between uncertain points and false detections. Extensive experiments on several public IRSTD benchmarks demonstrate that RefineFD consistently achieves better accuracy and lower computational cost than current methods. The source code is available athttps://github.com/duwl-SITP/RefineFD-master.