Lightweight aerial image object detection algorithm based on improved YOLOv5s.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection algorithms like YOLOv5 are crucial for image analysis but often require significant computational resources.
• Deploying these models on resource-constrained mobile and embedded devices presents a challenge due to their complexity.

Purpose of the Study:

• To develop a lightweight object detection algorithm optimized for aerial imagery on mobile and embedded devices.
• To improve the detection accuracy of small objects in aerial images while maintaining computational efficiency.

Main Methods:

• Proposed a lightweight aerial image object detection algorithm (LAI-YOLOv5s) by modifying YOLOv5s.
• Introduced a Deep Feature Map Cross Path Fusion Network (DFM-CPFN) for enhanced deep feature semantic information.
• Integrated a new module based on VoVNet to boost backbone network feature extraction.
• Applied principles from ShuffleNetV2 to reduce network weight without compromising accuracy.

Main Results:

• The LAI-YOLOv5s algorithm demonstrated an 8.3% higher mAP@0.5 detection accuracy on the VisDrone2019 dataset compared to the original YOLOv5s.
• Achieved lower computational cost and faster reasoning speed than other YOLOv5 series and YOLOv3 algorithms.
• Maintained high detection accuracy despite significant reduction in model complexity.

Conclusions:

• LAI-YOLOv5s is an effective lightweight object detection model for aerial imagery, suitable for mobile and embedded applications.
• The proposed enhancements significantly improve small object detection and overall performance.
• The algorithm offers a favorable balance between computational efficiency and detection accuracy.