Industrial Weld Defect Detection Based on Monocular Depth Estimation and Dual-Attention Point Cloud Network.

Area of Science:

• Industrial Quality Control
• Computer Vision
• Machine Learning

Background:

• Traditional 2D methods for weld defect detection suffer from illumination and texture interference.
• High-precision 3D laser scanning is expensive and impractical for large-scale industrial use.
• Accurate geometric defect identification is crucial for industrial quality control.

Purpose of the Study:

• To develop a cost-effective framework for reliable geometric weld defect detection.
• To address challenges in identifying complex missed weld defects with distinct spatial features.
• To enhance feature representation in point cloud networks for improved classification.

Main Methods:

• Utilized YOLOv8 for rapid region of interest extraction.
• Employed advanced monocular depth estimation to generate 3D pseudo-point clouds.
• Introduced a dual-attention enhanced point cloud classification network (DA-PointNet++) integrating dual-attention modules into PointNet++.

Main Results:

• Achieved 93.67% accuracy and 90.51% recall in a unified binary classification task.
• Effectively identified both normal welds and complex missed weld defects.
• Demonstrated significant reduction in false negative rates compared to PointConv, DGCNN, and Point Cloud Transformer.

Conclusions:

• The proposed framework offers a cost-effective solution for industrial automation in weld defect detection.
• DA-PointNet++ enhances feature representation for improved geometric defect identification.
• The approach provides reliable and low-cost geometric defect detection, overcoming limitations of existing methods.