Advanced defect recognition on scanning electron microscope images: a two-stage strategy based on deep convolutional neural networks for hotspot monitoring

We focus on detecting defects during hotspot monitoring using scanning electron microscope (SEM) images from the after-development inspection (ADI) phase in semiconductor manufacturing. A primary challenge lies in the significant imbalance between defective and defect-free samples, complicating binary classification. To address this, data augmentation and the synthetic minority over-sampling technique (SMOTE) are applied to generate additional samples for the minority class. A two-stage learning approach is then proposed: first, deep convolutional neural networks are trained on SMOTE-synthesized data to develop a pre-trained defect classification model. Subsequently, this model is fine-tuned with raw, augmented data to enhance its performance on targeted data. Extensive experiments demonstrate that this method improves efficiency, reduces computational costs, and outperforms traditional techniques. The results show that integrating SMOTE-based augmentation with a two-stage learning process mitigates data imbalance and improves defect detection accuracy. This proposed methodology offers a robust solution for automating SEM image inspections during hotspot monitoring, addressing the scalability challenges posed by increasing design complexity and emerging defect types in advanced semiconductor processes.