EMFE-YOLO: A Lightweight Small Object Detection Model for UAVs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Small object detection in Unmanned Aerial Vehicle (UAV) aerial imagery presents significant challenges, including low accuracy and complex backgrounds.
• Deploying large-parameter object detection models on resource-constrained UAVs is computationally prohibitive.

Purpose of the Study:

• To propose a lightweight small object detection model, EMFE-YOLO, designed for efficient deployment on UAVs.
• To enhance detection accuracy for small objects in complex aerial backgrounds while minimizing model parameters.

Main Methods:

• Developed EMFE-YOLO by improving the YOLOv8s architecture.
• Integrated the Enhanced Attention to Large-scale Features (EALF) structure to focus on large-scale features and improve small object detection.
• Incorporated the efficient multi-scale feature enhancement (EMFE) module for feature extraction and background interference mitigation.
• Utilized DySample in the neck of the network to optimize feature upsampling.

Main Results:

• EMFE-YOLO demonstrated significant improvements on the VisDrone2019-val dataset, with mAP50 increasing by 8.5% and mAP50:95 by 6.3% compared to YOLOv8s.
• The model achieved a substantial reduction in parameters, decreasing by 73% relative to YOLOv8s.
• Achieved a favorable balance between detection accuracy and computational efficiency.

Conclusions:

• EMFE-YOLO offers a viable solution for accurate and efficient small object detection in aerial imagery from UAVs.
• The proposed model's lightweight nature makes it suitable for deployment on UAVs with limited computational resources.