LSR-YOLO: A lightweight and fast model for retail products detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning object detection enhances retail product identification.
• Existing methods face challenges with high computational costs and slow speeds.
• Need for efficient models in smart cities and intelligent devices.

Purpose of the Study:

• Propose LSR-YOLO, a lightweight object detection framework based on YOLOv8n.
• Optimize the model for deployment in robots and intelligent devices.
• Improve inference speed and reduce computational load for real-time retail applications.

Main Methods:

• Developed LSR-YOLO with architectural optimizations, including the CSPHet-CBAM attention module.
• Implemented a channel pruning algorithm to reduce model redundancy.
• Evaluated performance on the Locount and COCO datasets.

Main Results:

• LSR-YOLO achieved 357.1 FPS inference speed on the Locount dataset.
• The model reached mAP50 of 72.2% and mAP50-95 of 47.8%.
• Demonstrated a 246.7 FPS increase over YOLOv8n with significantly fewer parameters and GFLOPs.

Conclusions:

• LSR-YOLO offers superior accuracy and computational efficiency for real-time retail object detection.
• The model's lightweight design and high speed make it suitable for resource-constrained devices.
• Validated generalization ability on the COCO dataset, confirming its practical applicability.