Design of Multi-scale Feature Fusion Algorithms for Infrared Small Target Detection

To address the challenges of low resolution, blurry imaging, and noise susceptibility in infrared small target detection, this paper proposes a multi-scale feature fusion infrared small target detection algorithm, YOLO-IRTD (YOLO-InfraRed Tiny Detect), based on YOLOv8n. The algorithm first reconstructs the feature fusion network of YOLOv8 and introduces a feature-focused diffusion pyramid network to enhance the network’s ability to extract small target features. It also designs an adaptive dynamic detection head that promotes information interaction between the classification and localization branches, improving detection accuracy while reducing model complexity. Finally, the SimAM attention mechanism is integrated into the Star Block to form the Star-SimAM module, which replaces the C2f module in the backbone network, further enhancing feature extraction capabilities while reducing the model's parameter count. Experimental results show that the YOLO-IRTD model reduces the parameter count by 20%, with a weight size of only 5 MB. It achieves a mAP of 90.4% on the HIT-UAV dataset, improving by 3.6% over the Base model and outperforming YOLO series models by 1%-4%. On the FLIR dataset, the mAP improves by 1.2%. The model demonstrates excellent generalization performance, and its overall metrics are among the best in current mainstream detection models.