Self-supervised real-noise fusion infrared small target detection network

Infrared small target detection (IRSTD) plays a crucial role in applications including, but not limited to, aerial surveillance, national defense, and disaster monitoring. However, existing detection models often suffer from limited generalization capability and localization accuracy due to data scarcity and significant variations in targets' scales and positions in real-world scenarios. To address these issues, this paper proposes a self-supervised real-noise fusion network for infrared small target detection. Specifically, we introduce a real-noise sampling and fusion mechanism that extracts authentic background noise from raw infrared imagery and integrates it into training samples, thereby producing more realistic synthetic data. Furthermore, a negative sample augmentation strategy is introduced to enrich negative samples with greater diversity and difficulty, alleviating the issue of limited dataset diversity. To further enhance the model's sensitivity to scale and location variations, a composite dynamic scale and location sensitive (DSLS) loss is designed. Additionally, a self-supervised optimization scheme is employed across multi-scale prediction outputs. Extensive experiments on two public infrared small target datasets validate the effectiveness of the proposed approach. Compared with the current leading method, DNANet, the proposed method achieves IoU improvements of 7.04% and 1.49% and detection rate increases of 0.91% and 0.96%, respectively. These results demonstrate the superior generalization and localization capability of the proposed framework under small-sample conditions, as well as its strong adaptability to variations in targets' scales and positions.