A real-time infrared small target detector based on an illumination-aware degradation filtering network

Infrared small-target detection remains challenging due to the extremely small size of targets, low target-background contrast, and severe interference from complex backgrounds and noise. To address these issues, this paper proposes an efficient real-time detection framework, termed IADF-YOLO, based on the YOLO paradigm. During training, a dual-backbone collaborative learning strategy is introduced to learn complementary feature representations. In addition, an illumination-aware degradation filtering network (IADF-Net) is designed to suppress illumination-induced degradation and noise-contaminated responses, while a content-aware cross-scale feature pyramid network (CACSFPN) is developed to enable efficient multi-scale feature fusion. Together, these components improve feature quality and multi-scale information utilization while preserving a lightweight architecture and real-time inference capability. Experiments on the ISA-TD and ISAD-6 datasets show that IADF-YOLO achieves a favorable balance among accuracy, model complexity, and inference speed. On ISA-TD, it achieves an mAP₅₀ of 0.936 and an mAP_50-95 of 0.740; on ISAD-6, it attains an mAP₅₀ of 0.949 with an F1 score of 0.92. With only 5.8M parameters and 11.9 GFLOPs, the proposed method consistently outperforms multiple state-of-the-art YOLO variants and other representative detectors. Further visualizations and qualitative comparisons show that IADF-YOLO maintains stable and reliable detection performance in practical scenarios involving complex backgrounds, long-range observation, and low-SNR conditions.