Top-Down Coarse-to-Fine Cascade Network for High-Precision Cluster Infrared Small Target Detection

Infrared small-target detection (IRSTD) holds a critical role in low-visibility and long-distance imaging scenarios, such as UAV tracking and maritime surveillance. However, cluster-IRSTD (CIRSTD) faces more prominent challenges: adjacent targets are prone to feature coupling, dim targets are easily submerged by background clutter, and cluster shapes vary dynamically. Owing to the constraint of independent single-target modeling, current deep-learning methods struggle to effectively handle dense cluster scenarios. Inspired by the human top-down visual attention mechanism, this paper proposes a coarse-to-fine cascaded detection network. First, an Adaptive Regional Attention (ARA) mechanism is tailored specifically for clusters, and a Coarse Cluster Extraction (CCE) module is further designed to extract the overall features of clusters. Subsequently, the Inner Fine Distinction (IFD) module seamlessly integrates the Gaussian and Scharr filters from model-driven approaches into the deep-learning framework, aiming to amplify the saliency of dim targets. It effectively solves the problems of dim target missed detection and adjacent target coupling in clusters. By synergistically integrating holistic cluster information and enhancing target saliency, the proposed Coarse-to-Fine Cascade IRSTD (C2IRSTD) significantly mitigates missed detections within clusters and reduces false alarms outside clusters. The experiments conducted on the DenseSIRST dataset have strongly demonstrated the superior performance of C2IRSTD in highly challenging dense-cluster scenarios. Meanwhile, its leading performance on the SIRST3 dataset in sparse scenarios fully highlights its excellent generalization ability. The code will be public athttps://github.com/wangtuntun/C2IRSTD.