Small target detection in UAV view based on improved YOLOv8 algorithm.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Detecting small targets in Unmanned Aerial Vehicle (UAV) imagery presents challenges like small object size, dense distribution, and hardware limitations.
• Existing models often struggle with accuracy due to these constraints, necessitating specialized approaches.

Purpose of the Study:

• To propose an improved YOLOv8 algorithm tailored for small target detection from UAV viewpoints.
• To enhance detection accuracy and computational efficiency for UAV-based object recognition.

Main Methods:

• Modified the YOLOv8n model by incorporating a bi-directional feature pyramid network (BiFPN) for superior feature fusion.
• Replaced the C2f module with the C3Ghost module to reduce computational load while maintaining performance.
• Integrated a channel attention mechanism into the detection head and improved the Minimum Point Distance based IoU (MPDIoU) loss function with inner-IoU concepts.

Main Results:

• The enhanced YOLOv8n model demonstrated significant improvements on the VisDrone dataset.
• Achieved a 17.2% increase in mean Average Precision (mAP), 10.5% in precision (P), and 16.2% in recall (R).
• The modifications effectively addressed challenges related to small target detection in UAV imagery.

Conclusions:

• The proposed improved YOLOv8n algorithm offers a robust solution for small target detection in UAV applications.
• The integration of advanced modules and loss functions enhances detection capabilities for complex scenarios.
• This approach significantly boosts the performance of small object detection from the UAV perspective.