A Study on Class Imbalance Handling in the WM-811K Dataset

This paper proposes a data augmentation–based oversampling technique to mitigate severe class imbalance and improve the classification performance of wafer map defect patterns in semiconductor manufacturing. Using the publicly available WM-811k dataset, the training data—excluding the dominant None class—are augmented to match different proportions of the None class (0%, 25%, 50%, and 100%) in order to determine the most effective augmentation level. Three deep learning models— DenseNet121, EfficientNet-B2, and ResNet50—are employed for performance benchmarking. In addition to overall accuracy, evaluation metrics that account for class imbalance, namely the Geometric Mean (GM) and Weighted Geometric Mean (GM+), are used to assess model effectiveness. Experimental results show that ResNet50 yields the most stable performance in terms of GM+, and that 50% augmentation achieves comparable results to full (100%) augmentation, indicating that excessive augmentation may be unnecessary. These findings suggest that rotation-based oversampling tailored to the characteristics of wafer maps significantly enhances defect pattern classification and can support the development of machine-learning-based defect analysis systems in semiconductor manufacturing.