Satellite Imagery Convolution-Based Object Recognition

Abstract

This study addresses challenges in remote sensing object detection, proposing the RAST-YOLO algorithm that integrates Region Attention (RA) with Swin Transformer as the backbone. The method effectively handles issues like varied target scales, intricate backgrounds, and closely spaced small objects. Incorporating the C3D module optimizes the multi-scale problem for small objects, enhancing detection accuracy. Evaluations on DIOR and TGRS-HRRSD datasets demonstrate RAST-YOLO's state-of-the-art performance, surpassing baseline networks. Notably, the model achieves a substantial mean average precision (mAP) improvement on both datasets, showcasing its effectiveness and superiority. Furthermore, the lightweight structure ensures real-time detection, making RAST-YOLO a practical choice for efficient and robust remote sensing object detection. The study extends the analysis to other prominent models like YOLOv5s, YOLOv3, FasterRCNN, RetinaNet, YOLOv5x6, and YOLOv8. Notably, YOLOv5x6 stands out with an impressive 0.80% mAP or higher, suggesting its potential for further enhancing detection performance in remote sensing applications.