Topology-Preserving Elastic Deformation Augmentation Enables Robust Defect Detection in Data-Scarce Industrial...

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Automated defect detection in materials microscopy faces challenges due to limited data and imbalanced classes.
• Existing methods struggle with rare and complex defect types under data constraints.

Purpose of the Study:

• To develop a robust and transferable deep learning framework for multilabel classification of material defects.
• To address data scarcity and class imbalance in automated inspection tasks.

Main Methods:

• Integration of topology-preserving elastic deformation (TPED) for realistic data augmentation.
• Utilizing a lightweight, attention-guided ensemble learner for improved classification accuracy.
• Validation on directed self-assembly (DSA) block copolymer SEM images and a wood defect dataset.

Main Results:

• Achieved high micro- and macro-averaged F1-scores across diverse and rare defect types.
• Demonstrated a critical F1-score of 0.987 for defect-free vs. defective sample discrimination.
• Successfully transferred the framework to a distinct wood defect dataset, proving generalizability.

Conclusions:

• The proposed framework offers a scalable, high-precision solution for industrial inspection with limited labeled data.
• The open-source, no-code interface enhances accessibility for practical industrial applications.
• The method effectively overcomes data limitations in automated defect detection.