Swin-Transformer-Based YOLOv5 for Small-Object Detection in Remote Sensing Images

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Small object detection in remote sensing images faces challenges with low accuracy and imprecise localization.
• Existing methods often struggle to capture sufficient global context and extract fine-grained features crucial for small objects.

Purpose of the Study:

• To develop an improved detection architecture for enhanced small object identification in remote sensing imagery.
• To increase the mean average precision (mAP) for small object detection tasks.

Main Methods:

• An enhanced YOLOv5 architecture integrating Swin Transformer for improved feature extraction and global context retention.
• Incorporation of Complete-IOU (CIOU) for K-means clustering and anchor generation, and a weighted bidirectional feature pyramid network for feature fusion.
• Addition of an extra prediction head, new feature fusion layers, and Coordinate Attention (CA) mechanism specifically for small objects.

Main Results:

• The proposed method achieved a mean average precision (mAP) of 74.7% on the DOTA dataset.
• Demonstrated an 8.9% improvement in mAP compared to the standard YOLOv5 model.
• Effectively improved the accuracy of small-object detection in challenging remote sensing scenarios.

Conclusions:

• The novel architecture significantly enhances small object detection accuracy and positioning in remote sensing images.
• The integration of Swin Transformer, CIOU, and CA mechanisms contributes to superior feature representation and fusion.
• The method offers a promising solution for high-accuracy small object detection in aerial imagery.