A Strategic Framework for Evaluating Data Augmentation in Microscopic IC Image Analysis

Microscopic image analysis is one of the most reliable approaches for the hardware assurance of integrated circuits (ICs). Recently, the paradigm of IC image analysis has shifted from manual approaches to automated systems, often leveraging deep learning methods. Sufficient training data of a high quality is important for the success of deep learning methods, but it can be costly and time-consuming to obtain such training data. Data augmentation is a popular technique that can artificially increase the quantity and variation of the training data samples, and it has been demonstrated to greatly improve the performance of deep learning methods. However, it is challenging to determine the suitable augmentation techniques to apply for different IC image datasets. Image data augmentations could also be applied with varying magnitudes, which further complicates the selection of augmentation techniques. In this paper, we propose a strategic framework to evaluate the performance of several data augmentation techniques for IC image analysis. We present a generic augmentation technique that combines the top performing individual augmentation techniques. We identify a reliable data augmentation strategy for IC image analysis through extensive experiments on seven datasets of microscopic IC images.