Proceedings Article | 27 March 2022
KEYWORDS: Target detection, Remote sensing, Neck, Image enhancement, Visualization, Satellite imaging, Image segmentation, Superposition, Satellites
Deep learning has achieved great success in the field of general target detection. However, remote sensing target detection have different characteristics from general image target detection in many aspects. Due to the high orbit height (usually over 500 km), the target usually only occupies a few pixels, which makes it difficult to detect. In addition, targets such as ships, trucks, etc. have a large aspect ratio, so traditional anchor designs are difficult to cover completely, resulting in failure of target detection because the intersection ratio of IOU is less than the typical value of 0.5. In addition, such targets are usually densely arranged when they are docked. The random imaging angle of satellite causes the traditional horizontal bounding box detection algorithm to be unable to accurately locate the target position. As a result, the deep learning training process is mixed with background information and cannot effectively learn the target features. Therefore, the oriented bounding box is used for labeling and training the remote sensing image. As a single-stage target detection algorithm and after several versions of the update, the detection accuracy and detection speed of the Yolo series algorithm have been greatly improved compared with the previous version. Based on the latest version of the Yolov5, the algorithm is modified to make it possible to support rotating target detection. Through the Gao-Fen competition of fine-grained classification of targets such as airplanes, cars, and ships (such as Boeing737, Boeing747, Boeing777, Boeing787, etc.), the average classification mAP value reaches 22.9235. For comparison, the circumscribed rectangle of the oriented bounding box is taken as the horizontal box which is used for another object detection called YOLOX training process, and the horizontal box target detection mAP value is 9.4679. The average classification mAP value of the modified Yolov5 algorithm is 13.4556 higher than the YOLOX-based horizontal box target detection mAP value. The rotating box training data is trained with another open source code called RotationDetection provided by Yangxue, The average classification mAP value obtained is 21.0232, which is lower than that of the modified Yolov5 algorithm by 1.9003. Therefore, the modified Yolov5 algorithm and the oriented bounding box have certain advantages, which can meet the needs of fine-grained rotating target detection in remote sensing images.
