WeldLight: A Lightweight Weld Classification and Feature Point Extraction Model for Weld Seam Tracking.

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Traditional vision-based weld tracking systems struggle with intense image noise and high computational costs.
• Accurate weld seam classification and feature point positioning are crucial for automated welding processes.

Purpose of the Study:

• To develop a lightweight and noise-resistant convolutional neural network (CNN) for precise weld seam feature point classification and positioning.
• To improve the adaptability and stability of vision-based weld tracking systems in noisy environments.

Main Methods:

• Proposed WeldLight, a one-stage lightweight CNN incorporating an online data augmentation method for noise adaptability.
• Implemented an attention module to filter noise-corrupted features, enhancing system stability.
• Utilized single-line structured light vision for seam feature point detection.

Main Results:

• Achieved an F1-score of 0.9668 for seam classification on an adjusted test set.
• Demonstrated low mean absolute positioning errors: 1.639 pixels (low-noise) and 1.736 pixels (high-noise).
• Inference time of 29.32 ms on a CPU platform, meeting real-time requirements.

Conclusions:

• WeldLight offers a robust and efficient solution for vision-based weld tracking, effectively handling intense image noise.
• The proposed network meets real-time performance demands for industrial seam tracking applications.
• WeldLight enhances precision and stability in classifying and positioning welding seam feature points.