Target Detection of Remote Sensing Image Based on an Improved YOLOv5

To solve the problems of low target detection accuracy caused by complex background of remote sensing images, dense distribution of small targets, various target scales and easily affected by environmental factors, this letter proposes a new target detection algorithm based on an improved YOLOv5. First, the coordinate attention mechanism is added in the last layer of the YOLOv5 backbone network. The coordinate attention can effectively capture the spatial relationships between pixels, thus enhancing the location-awareness of the model to suppress the interference of redundant information. Second, the small target detection layer is added to the neck in the original structure, and three scale feature detection heads are increased to four, and the new detection heads are detected on a 160×160 feature map, which reduces the receptive field and can enhance the multi-scale target detection performance of the algorithm to solve the phenomenon that small targets are closely distributed and easy to miss detection. Finally, experiments and validation were conducted in the DOTA dataset. Experiment results show that our improved algorithm can effectively improve the accuracy of remote sensing images target detection. The mean Average Precision (mAP) of the improved YOLOv5 algorithm reached 70.2%, which is 2.8% higher than the original YOLOv5.