HATSC-YOLOv10:Improved YOLOv10 for Satellite Remote Sensing Images of Small Object Detection

When dealing with small object detection in satellite remote sensing images, the state-of-the-art object detection model YOLOv10 often suffers from the difficulty of extracting enough local detail features of small objects. In this paper, a novel and improved YOLOv10 network based on hybrid attention transformer (HAT) as well as spatial and channel reconstruction convolution (SCconv) is proposed, which is named HATSC-YOLOv10. A brand new HATSC module, which takes full use of HAT and SCconv to extract and refine more efficient high-level features, is designed and integrated with original YOLOv10 network so as to strengthen its capability of feature extraction towards small objects. What's more, the normalized Gaussian Wasserstein Distance (NWD) metric is used to calculate the similarity between two bounding boxes instead of the complete intersection over union (CIoU) metric used in the loss function of original YOLOv10, in order to make the whole network be insensitive to the positional deviation of small objects. Experimental results on DOTA remote sensing dataset show satisfactory performance of the proposed HATSC-YOLOv10 on small object detection in satellite remote sensing images. Compared with original YOLOv10 network, HATSC-YOLOv10 shows an increase of 2.61%, 2.09%, 2.3%, and 1.68% in precision, recall, mAP50, and mAP50-95 respectively.