Unsupervised domain adaptation for IC image segmentation with structural constraint and pseudo supervision

Abstract

Integrated circuit (IC) image segmentation is crucial for functional verification and trustworthiness evaluation of ICs manufactured in the globalized supply chain. Conventional supervised deep learning models for IC image segmentation face significant challenges of domain shift when applied across IC layers. To address this, we propose a domain adaptation framework with Structural Constraint and Pseudo Supervision (SCPS) for improving segmentation performance on target datasets collected from different IC layers. Our proposed SCPS first leverages CycleGAN to synthesize target dataset with input masks from source dataset, where a constraint is imposed onto the structural patterns of circuit elements in synthetic target images and source masks to improve their structural consistency. It further utilizes unlabeled real target images through domain mixing and image−/feature-level augmentation with pseudo supervision during training. With experiments on target datasets collected from two different IC chip layers, our proposed SCPS outperforms existing methods in the accuracy of circuit connections retrieved from IC image segmentation, while maintaining comparable performance in terms of commonly used segmentation metrics.