A Multi-Scale Frequency Feature Fusion Model for Infrared Small Targets

In the field of computer vision, infrared small target detection generally faces tough challenges including tiny target sizes, blurry contour edges, strong background clutter, and low image signal-to-noise ratio (SNR), which easily lead to missed detections and false alarms during detection. This paper proposes a multi-scale frequency feature fusion model named YOLOv8-MSF based on YOLOv8n for infrared small targets. Firstly, pinwheel convolution is introduced to reconstruct the bottleneck structure, and the original C2f module is upgraded to the lightweight C2f-pinwheel convolution (C2f-P) module to strengthen the feature extraction capability of the central receptive field while cutting down model parameters and computational overhead. Secondly, a Multi-Scale Frequency Feature Fusion Module (MSFM) is designed to dynamically fuse multi-scale high-frequency and low-frequency features under the guidance of the attention mechanism, which effectively restores blurry target edges and suppresses complex background noise. Finally, the WIoUv3 loss function is adopted to boost detection stability. Multiple comparative experiments are carried out on the public HIT-UAV and FLIR infrared datasets. Compared with the baseline YOLOv8n, YOLOv8-MSF achieves a 10.7 percentage point (pp) improvement in precision and a 3.9 pp improvement in mAP50, with parameters reduced by 13.3% and GFLOPs cut by 13.5%, while maintaining real-time inference speed. The proposed model delivers outstanding accuracy, lightweight performance, and generalization capacity, making it applicable to practical infrared weak target detection scenarios.