Surface defect detection and classification with deep learning methods in semiconductor industry

Abstract

In the semiconductor sector, which is vital for producing microelectronic devices and sensors, maintaining high quality is imperative due to the complexity of the manufacturing processes. Despite advancements, manual inspection remains integral, primarily due to the sophisticated nature of defects. However, this method is both labor-intensive and limited in accuracy and efficiency. This research implements a deep learning method, specifically using Convolutional Neural Networks (CNNs), to automate the classification and detection of defects. A novel methodology that employs the Faster R-CNN model, combined with a sliding window approach, has been implemented. This methodology encompasses four unique architectures: Resnet50, MobileNet, VGGNet, and ViT-based. The sliding window technique proved beneficial in reducing computational requirements during the training phase with high-resolution imagery. The findings revealed that the highest binary classification was achieved by MobileNet with an F1-score of 0.91. In contrast, Resnet50 paired with ViT excelled in the IoU metric, indicating superior defect localization capabilities. While the potential of deep learning in this sector is evident from the research, further enhancements are necessary. The complexity of defect structures and the varied sensitivity of semiconductor surfaces still require a final human review for optimal defect verification.