Dual Attention-Based Industrial Surface Defect Detection with Consistency Loss.

Area of Science:

• Industrial Engineering
• Computer Science
• Materials Science

Background:

• Surface defects are common in industrial production, leading to unqualified products and impacting quality control.
• Detecting diverse surface defects is challenging due to varying manifestations and the rarity of defective samples.
• Semi-supervised anomaly detection is a suitable approach for surface defect detection problems.

Purpose of the Study:

• To propose an effective anomaly detection method for industrial surface defect detection.
• To enhance the accuracy and robustness of defect identification in manufacturing.
• To address the challenges of diverse defect types and limited sample availability.

Main Methods:

• Developed an anomaly detection method based on dual attention (channel and pixel attention) for robust normal image reconstruction.
• Introduced a consistency loss function to leverage multi-modal image differences for improved anomaly detection performance.
• Utilized semi-supervised learning to handle the scarcity of defective samples.

Main Results:

• The proposed dual attention mechanism enhanced the reconstruction of normal images, aiding in defect differentiation.
• The consistency loss function effectively improved the overall performance of the anomaly detection system.
• Experimental results demonstrated superior performance compared to existing methods on the Magnetic Tile and MVTec AD datasets.

Conclusions:

• The proposed dual attention and consistency loss-based anomaly detection method is highly effective for industrial surface defect detection.
• This approach offers a robust solution for identifying diverse and rare surface defects in manufacturing.
• The method shows significant potential for improving industrial product quality control and production efficiency.