SRE-YOLOv8: An Improved UAV Object Detection Model Utilizing Swin Transformer and RE-FPN.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unmanned aerial vehicle (UAV) imagery presents challenges for object detection due to object size variation and limited features.
• Existing object detection algorithms struggle with accuracy in complex aerial scenes.

Purpose of the Study:

• To enhance the detection accuracy of objects in UAV imagery.
• To address limitations in detecting small and low-resolution targets within complex backgrounds.

Main Methods:

• The study proposes SRE-YOLOv8, an enhanced YOLOv8 algorithm incorporating a Swin Transformer for global context and a lightweight residual feature pyramid network (RE-FPN).
• Key components include a Residual Feature Augmentation (RFA) module, ECA attention, a small object detection (SOD) layer, and a Dynamic Head with multiple attention mechanisms.
• The Swin Transformer optimizes feature extraction by preserving global context via self-attention.

Main Results:

• Experimental evaluation on the VisDrone2021 dataset demonstrated a significant improvement in detection accuracy.
• The SRE-YOLOv8 method achieved a 9.2% enhancement compared to the original YOLOv8 algorithm.
• The integrated modules effectively improved the emphasis on critical features and small object recognition.

Conclusions:

• SRE-YOLOv8 offers a robust solution for improving object detection accuracy in UAV imagery.
• The method's enhancements, particularly for small and complex targets, show its potential for real-world aerial surveillance and analysis.
• The integration of Swin Transformer and RE-FPN provides a powerful framework for advanced computer vision tasks.