Unsupervised Domain Adaptation with Pseudo Shape Supervision for IC Image Segmentation

Deep learning (DL) techniques have achieved excellent results for IC image segmentation, a critical task in hardware assurance, but they require a large amount of labeled training data to perform well. Due to the domain shift problem of DL techniques, the data collection and data labeling process has to be repeated on new IC image datasets which is extremely time-consuming. Domain adaptation is a promising approach that aims to tackle the domain shift problem by implementing models that can be trained on an existing source dataset and then applied to a new target dataset effectively. However, the reported domain adaptation techniques do not fully utilize the unlabeled images for training or are prone to model collapse when training on unlabeled images. To address these challenges, we propose pseudo shape supervision (PSS), a domain adaptation framework for IC image segmentation that effectively leverages unlabeled target images for training whilst avoiding model collapse. Within our PSS, we propose a novel shape consistency loss for supervision on unlabeled images, by utilizing weak pseudo-labels that are generated by thresholding. Cross domain mixing is performed between the unlabeled target images and the synthetic images to reduce the domain gap. Our experimental results demonstrate that our proposed PSS outperforms the reported techniques on IC image datasets and our ablation studies show the importance of our novel shape consistency loss.