The current research landscape in defect detection and classification across various manufacturing processes is witnessing significant advancements, particularly in the application of deep learning and machine learning techniques. A notable trend is the shift towards more data-efficient and self-supervised learning frameworks, which are proving to be highly effective in scenarios where labeled data is scarce or costly to obtain. For instance, the use of Vision Transformers (ViT) and Masked Autoencoders (MAE) in self-supervised learning for in-situ melt pool characterization in Directed Energy Deposition (DED) demonstrates a promising direction for automated quality control with minimal labeled data. Similarly, the integration of autoencoder-based data augmentation with convolutional neural networks (CNNs) for wafer map defect classification showcases enhanced generalization capabilities and robustness against class imbalance and noisy data. These approaches not only improve classification accuracy but also reduce the dependency on extensive labeled datasets, making them more feasible for real-world applications. Additionally, novel sequential learning frameworks like SL-RF+ for melt pool defect classification in Laser Powder Bed Fusion (L-PBF) are emerging as powerful tools for maximizing data efficiency and model accuracy in data-scarce environments. These developments collectively underscore a transformative shift towards more intelligent, adaptive, and cost-effective solutions in defect detection and classification across diverse manufacturing processes.

Noteworthy papers include one that employs a Vision Transformer-based Masked Autoencoder for in-situ melt pool characterization, achieving high accuracy with minimal labeled data, and another that combines autoencoder-based data augmentation with CNNs for wafer map defect classification, significantly outperforming traditional methods in accuracy.