FSC-YOLOv5s: A target detection algorithm for aerial infrared scenes

Abstract

Aiming at the problems of complex background, low resolution, and target scale variation in infrared images observed from UAV, we propose an efficient infrared scene target detection algorithm (FSC-YOLOv5s), which can effectively improve the accuracy of infrared target detection in the perspective of UAV. Firstly, we propose the FC3 module to improve the backbone network of YOLOv5s, which has information fusion between the backbone networks. Then, the Swin Transformer module is added to the end of the feature extraction Spatial Pyramid Pooling Structure (SPPF) and the neck network feature fusion to fully extract the image feature information. Finally, on the basis of the original feature fusion, the shallow features were fused by downsampling to extract rich semantic information. Experimental results show that although the number of model parameters is increased slightly, the detection accuracy mAP reaches 92.3%, which is 2% higher than that of YOLOv5s, and the speed reaches 37.88FPS. It can be seen that FSC-YOLOv5s has faster convergence speed, higher detection accuracy, and good real-time performance, which is more conducive to practical applications.