关键词: YOLOv10n算法, 特征提取, 红外小目标检测, 轻量化

Abstract: Infrared small target detection has a wide range of applications in military fields such as infrared guidance and tracking systems, and is an important direction in infrared image processing. Due to the limitations of detection equipment and the lack of inherent information of infrared small targets, existing detection methods are difficult to meet actual performance requirements. In order to explore a lightweight and high-precision infrared small target detection model, this study designed a lightweight and efficient infrared small target detection model (L-YOLOv10n) based on YOLOv10n. The SCDown module in YOLOv10n was replaced by a lightweight downsampling module (L-SCDown), and the key features of infrared small targets were enhanced at a lower computational cost; the C2f module was replaced by a lightweight feature extraction module (L-C2f), which enhanced the edge information of small targets and extracted multi-scale features while reducing the computational cost. In view of the problem of small infrared targets with few pixels and imbalanced foreground and background, the Focal Loss and Focaler-IOU loss functions were introduced to make the model more focused on difficult-to-detect targets. Experimental results on the public datasets SIRST-V2 and NUDT-SIRST show that the detection performance and resource consumption of L-YOLOv10n are significantly better than those of detection-based models. For segmentation models, the detection performance of L-YOLOv10n is slightly lower than that of Transformer-based segmentation models, but the resource consumption of L-YOLOv10n is significantly better than that of other models; the generalization performance on the NUDT-SIRST dataset is also much higher than that of most infrared small target detection models. The research results show that the proposed model has achieved a balance between resource consumption and high-precision detection and has certain practicality. Our code is available at https://github.com/Zichen-Zhao01/L-YOLOv10n.

Key words: YOLOv10n algorithm, feature extraction, infrared small target detection, lightweight