Reproduce “LithoSeg: A Coarse-to-Fine Framework for High-Precision Lithography Segmentation” on OpenResearch

About this paper

Xinyu He, Botong Zhao, Bingbing Li, Shujing Lyu, Jiwei Shen, Yue Lu

LithoSeg presents a coarse-to-fine framework for high-precision lithography segmentation from SEM images, combining a human-in-the-loop bootstrapping approach with a finetuned vision foundation model for initial masks and a physics-informed, point-wise 1D regression module for pixel-level contour refinement. The method reduces annotation overhead and achieves metrology-grade accuracy, outperforming existing approaches across various semiconductor manufacturing scenarios.

Key insights

• Adapting large vision foundation models (like SAM) through an iterative, prompt-free finetuning and minimal human curation significantly reduces annotation costs and improves generalization for initial segmentation in specialized domains.
• Transforming complex 2D contour refinement into a series of 1D regression problems along normal vectors, guided by physics-informed scan sizes and feature alignment, can achieve metrology-grade precision efficiently.
• Centering 1D profiles along the brightest pixel within a scan range compensates for translational errors from coarse segmentation and ensures robust feature representation for the lightweight regressor.

Summary by alphaXiv.