Filtered Distillation from a Large Vision Teacher for Infrared Small Target Detection

Abstract

Infrared target detection supports the continuous observation of traffic participants and low-altitude targets across aerial and fixed-view imaging settings, particularly under weak or changing illumination. However, infrared targets are often small, weakly textured, and easily confused with thermal noise and background clutter. Large pretrained vision models offer strong representation and generalization capabilities, but their parameter counts and computational costs make direct deployment on edge platforms with limited resources impractical. To transfer these capabilities to lightweight models, this paper proposes knowledge distillation at the label level based on filtered teacher detections. A large teacher adapted to the infrared domain first generates candidate boxes. Candidate boxes are selected using confidence thresholds, class reliability, and spatial relationships with ground truth annotations. The retained teacher boxes and the original annotations jointly form the student training targets, while the student retains its standard detection loss and original inference structure. On an independent sequence-level test set, the YOLOv5n baseline obtains an mAP@0.5 of 0.2277 and an mAP@0.5:0.95 of 0.1070, whereas filtered label distillation obtains 0.2547 and 0.1153, respectively, over three matched seeds. The filtering rules and thresholds are fixed before this evaluation, and the fixed Epoch-30 checkpoint is used for every run. Both student architectures are deployed on RK3588.