LPO-YOLOv5s: A Lightweight Pouring Robot Object Detection Algorithm

Area of Science:

• Robotics and Automation
• Computer Vision
• Artificial Intelligence

Background:

• Traditional object detection methods struggle with accuracy in casting processes.
• Deep learning models for pouring hole detection are resource-intensive, hindering deployment on robots.
• Accurate identification and localization of pouring holes are critical for automated casting.

Purpose of the Study:

• To develop a lightweight object detection algorithm for identifying pouring holes in resource-constrained casting robots.
• To improve the efficiency and accuracy of pouring hole detection for automated casting processes.
• To enable the deployment of advanced computer vision models on embedded robotic systems.

Main Methods:

• A lightweight pouring robot hole detection algorithm, LPO-YOLOv5s, based on YOLOv5s was designed.
• MobileNetv3 was integrated as a feature extraction network to reduce model complexity and parameters.
• A depthwise separable information fusion module (DSIFM) and CARAFE for feature upsampling, along with a dynamic head (DyHead), were employed.

Main Results:

• LPO-YOLOv5s achieved a 45% reduction in parameter size and a 55% decrease in computational costs compared to YOLOv5s.
• The model exhibited a minimal drop of 0.1% in mean average precision (mAP).
• The final model size was reduced to 7.74 MB, meeting deployment requirements for pouring robots.

Conclusions:

• The proposed LPO-YOLOv5s algorithm effectively addresses the challenges of deploying deep learning models for pouring hole detection on resource-limited robots.
• The lightweight design ensures efficient performance without significant compromise on detection accuracy.
• This advancement facilitates the integration of intelligent vision systems in automated casting operations.