SwinRODNet Swin Transformer-Based Remote Sensing Object Detection Network

Abstract

Detecting objects via remote sensing in natural settings is extremely difficult, especially when dealing with small targets and complicated backdrops. To improve feature extraction and detection accuracy, this proposal presents the RAST-YOLO method, which combines the Region Attention (RA) mechanism with the dual Transformer backbone. To solve the multi-scale issue and improve small item detection, the C3D module is used to combine deep and shallow semantic information. The algorithm's exceptional resilience, accuracy, and efficiency are demonstrated by extensive testing on the DIOR and TGRS-HRRSD datasets. Compared to baseline networks, RAST-YOLO shows a notable improvement in mean average precision (mAP) on both datasets. Additionally, research using methods like YOLOv5x6 and YOLOv8 reveals promising results, with YOLOv5x6 achieving a significant mAP enhancement of over 0.80%, highlighting its suitability for advanced remote sensing object detection applications. The proposed algorithm's ability to handle complex backgrounds and small-scale targets effectively makes it a valuable tool for various remote sensing applications, including environmental monitoring, resource exploration, and intelligent navigation.