An improved lightweight object detection algorithm for YOLOv5.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning-based object detection is widely used but faces challenges in mobile/edge deployment due to model size and computational cost.
• Lightweight models offer reduced parameters and computational needs but often sacrifice detection accuracy.

Purpose of the Study:

• To propose an improved lightweight object detection model based on YOLOv5s.
• To achieve higher detection accuracy with fewer parameters for efficient deployment on resource-constrained devices.

Main Methods:

• Integrated Ghost modules into the C3 structure to reduce parameters and speed up inference.
• Incorporated coordinate attention (CA) mechanism in the neck to enhance feature focus and accuracy.
• Designed a Simplified Spatial Pyramid Pooling-Fast (SimSPPF) module to improve model stability and reduce training time.

Main Results:

• Achieved a 28% reduction in model parameters compared to the original YOLOv5s.
• Increased mean average precision (mAP) by 3.1%, 1.1%, and 1.8% across three diverse datasets.
• Outperformed state-of-the-art lightweight models like YOLOv7tiny and YOLOv8n in mAP.

Conclusions:

• The proposed lightweight object detection model effectively balances reduced parameters with improved accuracy.
• Offers a valuable reference for deploying accurate and efficient object detection on mobile and edge devices.