Aerial image object detection based on improved YOLOv5

Object detection from aerial images has been a popular research field recently. The task, i.e., detecting objects from aerial images, is difficult due to the relatively large image dimensions and various directions of objects. In this paper, to improve the detection ability of small targets, we use mosaic for data enhancement during data preprocessing. A multi-scale test is performed on the test set to make the test result more accurate. In addition, an improved YOLOv5 algorithm, which introduces the circular smoothing label (CSL) angle processing method into YOLOv5, is proposed to detect rotating objects from aerial images. Experimental results on the DOTA-v2.0 dataset show that the proposed method is effective.