Long-tailed detection and classification of wafer defects from scanning electron microscope images robust to diverse image backgrounds and defect scales

Abstract

In semiconductor engineering, high yield of wafers relies on accurate detection and classification of wafer defects. The dataset for detecting wafer defects presents three primary challenges: (i) different background types, (ii) variable image or defect scales, and (iii) imbalanced data with a long-tailed distribution of defect types. These challenges create significant limitations for traditional classification techniques. To address these issues, we propose a stratified framework called Wafer Detection and Classification (WaferDC), designed specifically for detecting and classifying wafer defects from scanning electron microscope (SEM) images. Our framework achieves high defect detection performance on SEM wafer images by utilizing a multi-cluster memory bank, which effectively handles the challenges of (i) variable background types and (ii) differing image or defect scales. Building on this robust detection, we propose Segmentation and Mix (SegMix), a novel defect augmentation technique based on anomaly heatmaps, which enhances the reliability of defect detection and classification in a (iii) long-tailed imbalanced environment. Finally, we pass defect-classified images through a parameter-efficient fine-tuning (PEFT)-based classifier (Shiet al., 2023) utilizing a vision transformer (ViT) architecture, further improving overall defect detection and classification performance. We rigorously tested WaferDC on a proprietary SEM wafer dataset and the public Describable Textures Dataset-Synthetic (DTD-Synthetic) and Magnetic Tile Defect (MTD) datasets. The results confirm the effectiveness of our method in improving defect detection and classification in wafer manufacturing. Our code is available at https://github.com/SpatialAILab/WaferDC.