Infrared intensity-aware feature learning for robust target detection in infrared images

Abstract

Infrared target detection is challenging because of strong background noise, weak foreground–background intensity contrast, blurred boundaries, and limited target pixels. Motivated by the challenges observed in infrared inspection scenarios, this paper proposes YOLOv11-IR-TSD, an infrared-adaptive detector for improving feature representation and localization accuracy. In the backbone, a lightweight DnCNN-Lite module is introduced to suppress infrared imaging noise, while the original spatial pyramid pooling structure is replaced with an infrared-intensity-aware spatial pyramid pooling–fast (IIA-SPPF) module to enhance multi-scale feature representation. In the neck, a DINOv2-assisted self-supervised contrastive guidance mechanism and an infrared-intensity-gradient-weighted fusion strategy are employed to improve feature discrimination under complex backgrounds. In addition, a feature alignment constraint, an infrared-intensity-adaptive activation function, and a lightweight bounding-box refinement module are incorporated into the detection head to improve cross-stage feature consistency and localization accuracy. Experiments on the public FLIR and LLVIP infrared detection benchmarks show that YOLOv11-IR-TSD achieves mAP@0.5 values of 91.0% and 89.5%, respectively, while maintaining inference speeds of 64 and 62 FPS. Compared with existing infrared detection approaches, the proposed detector achieves improved detection accuracy while maintaining efficient inference, and exhibits enhanced robustness under noise interference, weak intensity contrast, and complex background conditions. Additional analyses under different target scales and contrast conditions further indicate the effectiveness of the proposed feature-learning strategy for challenging infrared scenes. Further evaluation on dedicated real-world infrared insulator datasets is required to investigate its applicability to abnormal-heating inspection.