Laser weld spot detection based on YOLO-weld.

Area of Science:

• Industrial Manufacturing
• Computer Vision
• Machine Learning

Background:

• Laser weld point detection is vital but challenged by limited, unevenly distributed, and irregularly shaped samples.
• Existing methods struggle with data scarcity and diverse object appearances.

Purpose of the Study:

• To develop an innovative, lightweight model (YOLO-Weld) for enhanced laser weld point detection accuracy and efficiency.
• To address data imbalance and improve bounding box regression in challenging industrial scenarios.

Main Methods:

• Employed targeted data augmentation for minority classes and introduced Diverse Class Normalization Loss (DCNLoss) to prioritize tail data.
• Developed Adaptive Hierarchical Intersection over Union Loss (AHIoU Loss) to focus on moderate IoU samples, speeding up regression.
• Proposed a lightweight multi-scale feature processing module, MSBCSPELAN, to optimize feature handling and reduce model size.

Main Results:

• YOLO-Weld significantly improved detection accuracy, with mAP@50 increasing by 15.6% and mAP@50:95 by 15.8%.
• The model achieved a 4.3% increase in precision, a 22.2% rise in recall, and a 15.1% increase in F1 score.
• Parameter count was reduced by 0.4 M, and GFLOPS decreased by 1.1, indicating a more efficient model.

Conclusions:

• YOLO-Weld offers a superior solution for laser weld point detection, overcoming data limitations and improving performance.
• The proposed model demonstrates enhanced accuracy, efficiency, and robustness for industrial applications.