Unpaired Synthesis of IC Scanning Electron Microscopy Images with Structural Constraints

The analysis of Integrated Circuit (IC) Scanning Electron Microscopy (SEM) images plays a crucial role in the reliability and authenticity investigation of modern ICs. It pertains to the task of predicting masks, which indicate the pixel-level position of different circuit elements, from IC SEM images, and it is now usually achieved through deep learning models that need to be trained or finetuned over a considerably large labeled dataset of IC SEM images. Furthermore, such labeled dataset is usually not publicly available and is costly to obtain. Recently, unpaired synthesis of IC SEM images, which aims to translate masks indicating IC circuit elements into IC SEM images with corresponding circuit elements, has made it possible for semi-supervised learning models to perform comparably well in IC image analysis but with lower dependency on labeled data. However, existing methods for unpaired synthesis of IC SEM images suffer from the low consistency of circuit element patterns between masks and synthetic images. In this paper, we propose a Structural Constrained Cycle-Consistent Adversarial Network (SC-CycleGAN), that leverages on CycleGAN and novel structural constraints to synthesize IC SEM images with improved consistency of circuit element patterns. Experiments on real masks and IC SEM images have demonstrated the superior performance of our proposed SC-CycleGAN.