Active Learning for Photomask Defect Classification

Sammanfattning

Supervised deep learning models have demonstrated strong performance for image classification tasks, but their success typically depends on access to large amounts of labelled training data. In photomask defect classification, obtaining such labels can be expensive and time-consuming since annotation requires specialized domain expertise. This thesis investigates active learning for deep learning-based photomask defect classification under class imbalance. Several acquisition strategies, ranging from established active learning methods to approaches specifically designed for class imbalanced data, are evaluated on a proprietary photomask defect dataset from Mycronic together with the public PathMNIST benchmark dataset. The results show that active learning consistently improves label efficiency compared to random sampling, reducing the number of labelled samples required to achieve a given level of classification performance. Furthermore, methods specifically designed to better account for class imbalance generally achieve the strongest overall performance and result in substantially more minority-class samples being acquired. The findings additionally indicate that practical design choices, such as query batch size and initial labelled budget, strongly influence performance when only a limited number of labelled samples are available.