CAST-DETR: Context-aware spatial-statistical transformer for UAV infrared small target detection

Abstract High-altitude aerial infrared (IR) object detection has become crucial for intelligent perception as unmanned aerial vehicles (UAVs) are increasingly used in all-weather missions, including night reconnaissance and border patrol.However, constrained by fluctuations in UAV flight altitudes and long-range imaging conditions, infrared small targets typically occupy extremely sparse pixel areas and suffer from a severe lack of texture information under low-contrast backgrounds. Consequently, practical applications of existing methods face severe challenges, including the vulnerability of target features to being submerged, the loss of multi-scale information, and intense background clutter interference. To address these issues, this paper proposes a Context-Aware and Statistical feature enhanced Transformer detection framework, termed CAST-DETR, tailored for UAV-based infrared small target detection. This framework establishes a synergistic enhancement mechanism spanning feature perception, multi-scale fusion, and global semantic interaction to comprehensively improve the representation capability of weak targets. Specifically, a Context-Aware Feature Enhancement Module (CAFEM) is first designed to jointly model local details, surrounding context, and global semantic information, thereby reinforcing the discriminative feature representations of infrared small targets under low SNR conditions. Secondly, a Collaborative Multi-scale Reconstruction Neck (CMR-Neck) is constructed, which integrates efficient upsampling and spatial cyclic displacement reorganization mechanisms to facilitate the cross-scale propagation and physical fidelity of fine-grained spatial details. Finally, a Token Self-supervised Inverse Feature Interaction (TSIFI) module based on second-order statistical features is introduced to enhance global semantic interaction via statistical response modeling, allowing the network to focus selectively on sparse, weak target regions while suppressing non-target background clutter. Comprehensive evaluations on both the HIT-UAV and AIR-STD datasets demonstrate that the CAST-DETR framework exhibits superior performance in infrared target detection. Specifically, on the HIT-UAV dataset, CAST-DETR achieves a mAP50 of 82.8%, yielding a 4.8 percentage point improvement over the baseline RT-DETR-r18, while achieving a mAP50 of 92.5% on the AIR-STD dataset.