YOLO-FSD: An Improved Target Detection Algorithm on Remote-Sensing Images

Target detection is one of the research hotspots in remote-sensing image information extraction, which has a wide range of application prospects. Compared with natural images, remote-sensing images are characterized by many dense small targets, large changes in target scale, and fuzzy target appearance under complex backgrounds. Therefore, this article proposes the YOLO-FSD remote-sensing image target detection algorithm. First, to address the problem of fuzzy target appearance under complex background, a Swin-CSP structure is introduced in each layer of the network, which enhances the model’s discriminative ability by learning the target’s local and global attributes. Then, the Faster-SPPCSPC module is designed to efficiently solve the problem of large target-scale variation by improving the feature pyramid. Finally, a new DWC-head is utilized to reduce the prediction bias due to tiny and dense targets and to improve the model’s localization and classification ability in complex backgrounds. Our extensive experiments show that our proposed YOLO-FSD algorithm improves average precision (AP) and mean AP (mAP) by 3.9% and 3.3%, respectively, on the VisDrone dataset compared to the original YOLOv7 algorithm. The AP and mAP are improved by 3.6% and 4.2% on the DOTA dataset, and the inference speed reaches 15.3 ms, which has better detection accuracy and inference speed.