Industrial Product Surface Anomaly Detection with Realistic Synthetic Anomalies Based on Defect Map Prediction

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface anomalies in industrial products degrade quality and pose safety risks.
• Limited defect samples hinder traditional deep learning for surface defect detection.
• Reconstruction-based anomaly detection is common but may reconstruct defects perfectly.

Purpose of the Study:

• To develop an effective surface defect detection algorithm for industrial products.
• To address the challenge of limited defect samples in deep learning models.
• To improve the accuracy and reliability of industrial quality control.

Main Methods:

• Proposed a reconstruction-based defect detection algorithm utilizing realistic synthetic anomalies for training.
• Implemented an auto-encoder image reconstruction network with deep feature consistency constraints.
• Introduced a defect separation network with a large receptive field.

Main Results:

• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) score of 99.70% on the MVTec anomaly detection dataset.
• Obtained an average precision (AP) score of 99.87% for defect detection.
• Demonstrated superior performance compared to recent defect detection algorithms.

Conclusions:

• The proposed method effectively handles the scarcity of defect samples in industrial surface defect detection.
• Utilizing synthetic anomalies and advanced network architectures enhances detection accuracy.
• The algorithm offers a robust solution for improving industrial product quality and production efficiency.