R-YOLO: A YOLO-Based Method for Arbitrary-Oriented Target Detection in High-Resolution Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Object detection in remote sensing is challenging due to varying resolutions, small/dense objects, and motion direction ambiguity.
• Existing methods struggle with the complexities of high-resolution aerial and satellite imagery.

Purpose of the Study:

• To improve the accuracy and robustness of object detection in challenging remote sensing image scenarios.
• To adapt the YOLOv5 architecture for enhanced feature extraction and rotation-invariant detection.

Main Methods:

• Proposed a modified YOLOv5 network incorporating the MS Transformer module for enhanced feature extraction.
• Integrated the Convolutional Block Attention Model (CBAM) to focus on dense object regions.
• Enhanced YOLOv5 with rotation angle detection and improved bounding box regression for oriented objects.
• Improved the focal loss function using a weighted combination of difficult sample mining techniques.

Main Results:

• Achieved an average accuracy of 77.01% on the DOTA dataset, surpassing previous algorithms.
• Demonstrated an 8.83% improvement in average detection accuracy compared to the standard YOLOv5.
• The enhanced algorithm showed superior detection performance in complex remote sensing environments.

Conclusions:

• The proposed modified YOLOv5 network effectively addresses challenges in remote sensing object detection.
• The integration of MS Transformer, CBAM, and rotation detection significantly boosts detection accuracy.
• The method offers a promising solution for accurate object detection in diverse remote sensing applications.