Abstract
The YOLO-B infrared target detection algorithm is proposed to address the problems of incomplete extraction of detailed features and missed and wrong detection of infrared targets by YOLOv5s. The algorithm improves the SPPF of YOLOv5s feature extraction network by proposing the CSPPF structure to increase the sensory field of the model. The Bifusion Neck structure is invoked to fuse the shallow location information with deep semantic information to enhance the feature extraction capability of the model. Taking fully into account the different information of concern for classification and localization, the efficient decoupled head is used as the prediction head of this algorithm, which reduces the latency while maintaining the accuracy. WIoUv3 loss is used as a bounding box regression loss function to reduce the harmful gradient generated by low-quality examples and reduce the competitiveness of high-quality anchor frames. Comparative experiments were conducted for each of the four improvement points, and the experimental results showed that each improvement point had the highest detection accuracy in the comparative experiments of the same category. All improvement points are fused in turn and ablation experiments are performed. The YOLO-B algorithm improves 1.9% in accuracy, 7.3% in recall, 3.8% in map_0.5, and 4.6% in map_0.5:0.95 compared to YOLOv5s. When compared with YOLOv7 and YOLOv8s, the proposed algorithm has better performance in terms of the number of parameters and detection accuracy.