Get started
Work

Efficient and Robust SEM Image Denoising for Wafer Defect Inspection

Year: 2025
Type: article
Abstract: Noise in scanning electron microscopy (SEM) often obscures details critical for accurate wafer defect inspection. Deep learning-based denoising methods have been widely used to address this problem, but they have two major limitations in SEM image denoising: lack of both efficient and powerful denoising methods, and poor generalization to image structures that are unseen during training. To this end, we propose Relaxed Noise2Noise with Input dropout (ReNIn), which includes components that address the above two issues. Firstly, our Relaxed Noise2Noise framework provides a much better trade-off between denoising performance and training data collection cost; namely, it shows nearly on-par denoising performance with much lower data collection cost than ordinary supervised learning-based methods. Secondly, we propose to apply the input dropout to boost generalization ability. It improves the performance of the images that are structurally different from the training images without compromising the performance of the normally structured images, thereby increasing the overall denoising performance. Consequently, our method supports downstream inspection tasks by reducing the failure rate in circle detection, which is a critical preprocessing step for circle-shaped product analysis. Overall, our ReNIn attains efficient training data collection cost with competitive denoising performance and enhances generalization capability across various structures.
DOI: 10.1093/mam/ozaf084
Language: English
Open Access status: hybrid
Estimated APC paid: $3,554
Funder: Samsung
OpenAlex ID: W4415544507
FWCI:
Cites: 8
Cited by:
Related to: 10