A Study of Defect Detection Techniques for Metallographic Images

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Metallography analyzes metal and alloy structures, aiding in identification, quality control, and defect characterization.
• Current metallographic defect detection relies heavily on human experts, presenting a significant challenge for automation.
• Deep learning, particularly convolutional neural networks, has shown promise in computer vision tasks.

Purpose of the Study:

• To develop a novel deep learning architecture for automated metallographic defect detection.
• To improve the accuracy and efficiency of identifying defects in metallographic images, especially small objects.
• To propose a practical system for real-world metallographic analysis applications.

Main Methods:

• A modified Residual Network (ResNet) architecture was developed, termed Multi-scale ResNet (M-ResNet).
• The M-ResNet incorporates multi-scale operations to enhance the detection of objects across various sizes.
• The model was trained and evaluated for its performance in recognizing defects in metallographic images.

Main Results:

• The proposed M-ResNet achieved a mean Average Precision (mAP) of 85.7% in recognition performance.
• This accuracy surpasses that of existing methods for metallographic defect detection.
• The multi-scale approach proved effective for detecting small objects within the images.

Conclusions:

• The developed M-ResNet offers a robust and accurate solution for automatic metallographic defect detection.
• This deep learning approach addresses the limitations of manual analysis in metallography.
• The system has significant potential as an application for enhancing metallographic analysis.