Improved YOLOv5 network method for remote sensing image objects detection

Aiming at the problems of missed detection and poor detection due to variable scale and complex background of remote sensing images, this paper proposes a network to enhance residuals and attention based on YOLOv5 model, named ERA-YOLO. The framework focuses on strengthening the feature representation capability of the backbone network. First, a Res2block module is designed in conjunction with the idea of res2net network structure to enable the network to represent multi-scale features at a finer level of granularity. Then, a coordinate attention module is inserted in the improved backbone network for improving the model's power to locate targets while suppressing the interference of background information. Experiments on a number of publicly accessible datasets demonstrate that the average detection accuracy is effectively improved compared to the original YOLOv5 and existing typical target detection models.