Hybrid Unsupervised Clustering for Pretext Distribution Learning in IC Image Analysis

Abstract

Delayered integrated circuit image analysis is an important step in hardware assurance, which is typically performed by automated approaches such as deep learning. The data dependent deep learning techniques require a diverse set of training data containing most of the variations in the delayered circuit images to perform well, which can be highly challenging to curate. In this paper, we present a hybrid unsupervised clustering method that aims to learn the distribution of newly acquired circuit image datasets, to aid the subsequent analysis flow. Our method consists of a deep learning-based feature extractor stage and a feature clustering stage, and we evaluate the performance of several feature extraction networks and clustering algorithms. Experimental results show that our method could obtain a promising normalized mutual information (NMI) score of 0.6095 on a dataset of delayered IC images taken of a manufactured Integrated Circuit (IC), and demonstrates excellent ability to retrieve visually similar images when provided with query images.