Thermal-aware UAV infrared small target detection integrating infrared domain knowledge

Infrared small target detection from unmanned aerial vehicles (UAVs) is essential for nighttime and low-visibility surveillance, search and rescue, and reconnaissance. Although recent prompting methods have improved infrared modality adaptation or mask-level infrared small target segmentation, multi-class UAV horizontal-box detection still requires the simultaneous resolution of three coupled problems: fine-grained semantic ambiguity in thermal imagery, loss of weak small-target evidence during multi-scale feature propagation, and scale-dependent localization errors under a practical deployment budget. We address this joint problem by organizing infrared domain knowledge around temperature-gradient and radiation-intensity cues and aligning it with visual features through Thermal-aware Dual-branch Semantic Alignment (TDSA). A Three-path Thermal-aware Scale-adaptive Hybrid Feature Fusion Block (TS-HFFB) preserves weak local evidence while incorporating global and cross-level context, and Scale-adaptive Target-Aware Reweighting (STAR) emphasizes difficult small-target localization. Training-only shape supervision further regularizes target boundaries without changing the deployed box-detection task. Controlled experiments show that the dual thermal prompt mapping outperforms class-name, random-prompt, and single-projection alternatives, while shape supervision improves AP50 by 0.98 points and mAP@[.5:.95] by 1.75 points. On the main benchmarks, the method achieves 86.8% mAP50 on HIT-UAV and 84.6% mAP50 on DroneVehicle, with 152 FPS and 10.8M parameters. The results support a problem-oriented thermal detection framework rather than a claim of being the first infrared prompting method.