Rotating object detection in remote-sensing environment

Deep learning models have become the mainstream algorithm for processing computer vision tasks. In tasks of object detection, detection box is usually set as a rectangular aligned with coordinate axis, so to achieve complete packaging object. However, when facing some objects large aspect ratio and angle, bounding must be enlarged, which makes contain amount useless background information. this study, different approach based on YOLOv5 adopted. By means, angle information dimension added at head, regression also same time boundary regression. Then loss calculated by combining ciou smoothl1, that obtained more closely suitable actual At time, original dataset tags are pre-processed calculate interest. The purpose these improvements realize angles in remote-sensing images, especially ratios, such ships, airplanes, automobiles. Compared traditional other state-of-the-art arbitrarily oriented model deep learning, experimental results show proposed method has unique effect detecting rotating objects.