RSI-YOLO: Object Detection Method for Remote Sensing Images Based on Improved YOLO.

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Object detection is crucial in computer vision, but challenging in remote sensing due to small object sizes and low resolutions.
• Existing deep learning models face difficulties in accurately detecting these objects.

Purpose of the Study:

• To propose an improved object detection algorithm, RSI-YOLO, based on YOLOv5 for enhanced remote sensing image analysis.
• To address the limitations of detecting small and low-resolution objects in remote sensing data.

Main Methods:

• Developed RSI-YOLO by integrating channel and spatial attention mechanisms into the YOLOv5 architecture.
• Improved the feature fusion structure from PANet to a weighted bidirectional feature pyramid.
• Incorporated a dedicated small object detection layer and optimized the loss function.

Main Results:

• RSI-YOLO demonstrated superior detection performance compared to the original YOLOv5 on multiple remote sensing datasets (e.g., DOTA, NWPU-VHR 10).
• The enhanced feature fusion and attention mechanisms significantly improved detection accuracy.
• The algorithm showed better performance than other classical object detection methods.

Conclusions:

• The proposed RSI-YOLO algorithm effectively enhances object detection in remote sensing images.
• The modifications to the YOLOv5 architecture lead to more robust and accurate detection of small objects.
• RSI-YOLO represents a significant advancement in applying deep learning for remote sensing image analysis.