Surface Defect Segmentation Algorithm of Steel Plate Based on Geometric Median Filter Pruning.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Deep learning models for defect detection often suffer from parameter redundancy, hindering deployment on resource-constrained factory embedded devices.
• Existing model compression techniques may not be optimal for structured pruning of defect segmentation networks.

Purpose of the Study:

• To analyze existing deep learning model compression algorithms.
• To propose a novel model pruning algorithm for efficient defect detection networks.
• To enable the application of defect detection models on factory embedded devices.

Main Methods:

• Analysis of various deep learning model compression algorithms.
• Development of a model pruning algorithm utilizing geometric median filtering.
• Application of structured pruning for defect segmentation detection networks.
• Experimental comparison and optimization of the proposed algorithm.

Main Results:

• Significant reduction in network parameters and computational effort.
• Effective pruning of defect detection algorithms for steel plate surfaces.
• Demonstrated feasibility for factory embedded device applications.

Conclusions:

• The proposed geometric median filtering-based pruning algorithm offers an effective solution for compressing deep learning defect detection models.
• This approach enhances the applicability of advanced defect detection in industrial settings with limited computational resources.