Data-Centric Model Development to Improve the CNN Classification of Defect Density SEM Images

Automatic classification in production environments is becoming increasingly important. It has an extremely positive effect on the efficiency of the workflow and provides objective and reproducible classification results. However, training models on real-world datasets is challenging, since they are most often small and highly imbalanced. This paper summarizes the data-centric model development approach we applied to train convolutional neural networks, through transfer learning, on a small and highly imbalanced defect density image dataset. In this data-centric approach, we iteratively assessed and modified our dataset to improve the classification performance during model training. With the final version of the dataset, we trained a model achieving a validation accuracy of 92.7%.