Table 1 Summary of the state-of-the-art defect classification papers

Machine Learning [18,19,20]

Surface defect classification; handcrafted feature extraction

Extracted features are sensitive to noise and variability, computationally intensive

Deep Learning [32, 33],

Lightweight architecture for defect detection; transfer of valuable defect features to student model, handle complex and varied defect patterns extracted using evolving convolutional architectures

Requiring large volumes of labeled examples, overfitting and poor generalization on rare classes, does not aid detection of rare and newer defects

TL-based methods [9, 28, 43, 44, 46]

Use of small dataset, saves training period; usage of prior knowledge towards target domain

If source and target tasks are not adequately representative, they become redundant; negative transfer

Adaptive Learning [50, 52, 53, 55, 57]

Defect detection with limited samples; process unlabeled data to obtain useful defect representations

Poor results for rare surface defects, computational complexity issue for real-time monitoring