Rotating target detection algorithm in remote sensing images based on improved YOLOv5s

Abstract

Aiming at the problem of low target detection accuracy and easily missed detection in remote sensing images, a rotating target detection method based on YOLOv5s is proposed with YOLOv5s as the basic framework. First of all, the circular smooth label CSL (Circular Smooth Label, CSL) is used to realize the conversion of angle regression prediction to classification prediction to overcome the problems of regression prediction angle periodicity and variable boundary and improve the detection accuracy. Secondly, the densely coded label (Densely Coded Label, DCL) replaces the sparsely coded label, dramatically reducing the thickness of the prediction layer and improving the training speed. Finally, in the C3 module of the original YOLOv5s feature extraction and feature fusion network, a lightweight channel attention mechanism is introduced into the module, thereby improving the network's local feature capture and fusion capabilities. The experimental results show that on the public DOTA remote sensing image dataset, the map results of the improved YOLOv5s_obb reach 0.78, respectively. Compared with the original YOLOv5s, this method can be improved by 1.7%, and compared with other typical remote sensing target detection methods, the accuracy has also been improved, which proves the effectiveness of the improved YOLOv5s_obb method.