From Anomaly Detection to Defect Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Inspection

Background:

• Automated defect detection is crucial for quality control in manufacturing.
• Existing methods often struggle to precisely localize diverse defect patterns.

Purpose of the Study:

• To develop a defect detection system that accurately identifies and highlights exact damaged areas in visual data.
• To propose a novel hybrid approach combining unsupervised and supervised learning for enhanced defect localization.

Main Methods:

• Utilized an autoencoder for anomaly detection by comparing original and reconstructed images.
• Applied DBSCAN clustering to group anomalies into regions of interest.
• Employed a pre-trained Xception network for supervised classification of detected defects.
• Combined these into an unsupervised-unsupervised-supervised (U2S-CNN) methodology.

Main Results:

• The system successfully identified 177 regions, with 108 correctly pinpointing 205 occurring damaged areas.
• Achieved accurate localization of defects, demonstrating the system's focus on exact damaged areas.
• Showcased the capability to detect a wide range of defect patterns.

Conclusions:

• The proposed U2S-CNN system provides a viable proof of concept for precise defect area detection.
• This approach offers a potential alternative to established methods like YOLO, autoencoders, and transformers for industrial inspection.
• Highlights the effectiveness of combining different machine learning paradigms for complex visual tasks.