An optimized lightweight real-time detection network model for IoT embedded devices.

Area of Science:

• Computer Vision
• Embedded Systems
• Deep Learning

Background:

• Internet of Things (IoT) devices require efficient real-time target detection for applications like intelligent manufacturing and autonomous driving.
• Deploying advanced models like YOLOv8 on resource-constrained IoT embedded devices presents significant computational challenges.

Purpose of the Study:

• To develop and deploy an optimized, lightweight real-time detection network model (FRYOLO) suitable for IoT embedded devices.
• To address the limitations of deploying high-performance deep learning models on devices with limited computing resources.

Main Methods:

• Proposed and deployed FRYOLO, an optimized lightweight real-time detection network.
• Evaluated FRYOLO's performance through a case study on real-time fresh and defective fruit detection in a production line.

Main Results:

• FRYOLO achieved 84.7% recall, 92.5% precision, and 89.0% mean Average Precision (mAP).
• The model demonstrated a detection frame rate of up to 33 Frames Per Second (FPS), meeting real-time requirements.
• FRYOLO exhibited low training cost and high detection performance for various fruit types and states.

Conclusions:

• FRYOLO is a viable solution for real-time object detection on IoT embedded devices, overcoming resource limitations.
• The implemented intelligent production line system showcases FRYOLO's practical applicability in industrial IoT scenarios.
• FRYOLO provides robust technical support for efficient fruit production processes and demonstrates effectiveness in real-world IoT applications.