Lightweight Corrosion Image Classification via Self-Training and Progressive Knowledge Distillation
TL;DR
A class-aware proportional screening strategy that expands a small, labelled dataset with 19,964 unlabeled images to create a more balanced training set and minimizes computational overhead without sacrificing accuracy, thereby offering a practical solution for real-world monitoring on resource-constrained devices.
Abstract
Automated corrosion classification is crucial for industrial inspection. However, existing methods face severe class imbalance and strict computational constraints on edge devices. This study proposes a stepwise self-training framework for six-category corrosion classification. Leveraging a Vision Transformer and XGBoost classifier, this study proposes a class-aware proportional screening strategy that expands a small, labelled dataset with 19,964 unlabeled images to create a more balanced training set. Progressive training and knowledge distillation are then integrated to train a lightweight MobileNetV2 student model. The optimal fine-tuned model achieves a peak classification accuracy of 87.52%, improving upon the baseline MobileNetV2 by over 5%. Grad-CAM analysis confirms accurate focus on corroded regions, while mask-based fine-tuning demonstrates robustness, maintaining over 83% accuracy under 10% visual interference. This approach minimizes computational overhead without sacrificing accuracy, thereby offering a practical solution for real-world monitoring on resource-constrained devices.