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

Automated defect detection in materials microscopy is frequently hindered by severe data scarcity and class imbalance. Here, we propose a practical and transferable deep learning framework designed to achieve robust multilabel classification under these constrained conditions. The method integrates a topology-preserving elastic deformation (TPED) strategy, which introduces realistic geometric variability while maintaining semantic consistency, with a lightweight, attention-guided ensemble learner. Validated on directed self-assembly (DSA) block copolymer SEM images, the framework effectively overcomes data set limitations, achieving high micro- and macro-averaged F1-scores across rare and complex defect types. Notably, the model attains an F1-score of 0.987 on the critical discrimination between defect-free and defective samples, approaching human-level precision. To confirm the framework's generalized utility beyond the semiconductor field, we demonstrate its successful transferability to a distinct wood defect data set. Encapsulated in an open-source, no-code graphical user interface, this work offers a scalable, high-precision solution for industrial inspection tasks where labeled data is limited.