Enhanced Vision-Based Quality Inspection: A Multiview Artificial Intelligence Framework for Defect Detection.

Area of Science:

• Industrial Automation
• Computer Vision
• Machine Learning

Background:

• Automated defect detection is crucial for industrial quality control.
• Subtle defects, such as scratches on metallic surfaces, pose significant detection challenges.
• Current single-view inspection methods often lack the necessary accuracy for complex defect identification.

Purpose of the Study:

• To investigate the effectiveness of multiview deep learning for enhanced defect detection.
• To compare early and late fusion methodologies in a multiview context.
• To propose and evaluate a novel early fusion architecture, MV-UNet, for improved accuracy.

Main Methods:

• Implementation and comparison of late fusion and early fusion deep learning approaches.
• Development of MV-UNet, an early fusion architecture utilizing a transformation block for feature alignment and aggregation.
• Experimental evaluation on a metallic plates dataset, comparing against single-view inspection.

Main Results:

• Both early and late fusion methods demonstrated improved detection accuracy compared to single-view inspection.
• The proposed MV-UNet achieved the highest F1-score of 0.942.
• Adapted precision-recall metrics were introduced, offering more accurate evaluation for segmentation-based defect detection, especially for elongated scratches.

Conclusions:

• Multiview deep learning, particularly early fusion, offers significant advantages for industrial defect detection.
• MV-UNet provides a robust and scalable solution for enhancing the accuracy of automated quality control systems.
• The developed tailored metrics improve the evaluation of defect localization performance in challenging scenarios.