WireGC-Former: Surface defect segmentation method of steel wire ropes based on 3D point clouds

Area of Science:

• Computer Vision
• Materials Science
• Industrial Inspection

Background:

• 3D point clouds are crucial for industrial product surface analysis.
• Existing datasets lack sufficient quality and quantity for small industrial defect segmentation.
• Wire rope defects (wear, protrusion, broken wire) present challenges due to scale, distribution, and morphology.

Purpose of the Study:

• To create a high-quality dataset for wire rope defect point cloud segmentation.
• To develop an effective point cloud segmentation method for small industrial defects.
• To improve the detection and analysis of wire rope surface flaws.

Main Methods:

• Construction of the WireRope3D dataset with three defect types.
• Proposal of WireGC-Former (Wire Rope Graph Convolution Transformer) for point cloud segmentation.
• Integration of graph convolution and Transformer with specialized modules for feature perception and dependency modeling.

Main Results:

• WireGC-Former achieved a mean Intersection over Union (mIoU) of 80.32% on the WireRope3D dataset.
• Demonstrated effectiveness in segmenting small-scale, complex wire rope surface defects.
• Validated the method's capability in precise defect identification.

Conclusions:

• The WireGC-Former method offers a robust solution for small-target defect segmentation in industrial point clouds.
• The WireRope3D dataset provides a valuable resource for research and development in this area.
• The approach serves as a reference for semantic segmentation in complex industrial applications.