Single Image Printed Circuit Board Functional Similarity Clustering Using Vision Transformers

Previously, using stand-alone computer vision techniques for the analysis of printed circuit board function has been difficult due to visual ambiguity and shortages of labeled data. With advancements in self-supervised learning and vision transformer architectures, we demonstrate the potential for image based clustering and similarity measures. In this work, a method for the construction of a learned metric space well suited for printed circuit board image-based comparison is created using a combination of self-supervised learning and supervised fine tuning with a vision transformer with patch size 8. A set of 28,000 unlabeled and 500 labeled images are used to construct this space across 10 functionally labeled classes. Through the use of the DINO self-supervised method, informative components are primed for attention during pre-training, which ultimately allows the network to converge with this relatively small amount of labeled training data. Amongst other potential use cases discussed herein, the latent space after fine tuning is demonstrated to be useful for image retrieval as well as functional classification. Broad applications to circuit board trust and assurance are discussed as extensions to this work, including comparisons of devices and systems, automated reverse engineering for supply chain and security purposes, and an understanding of functional capabilities of an unknown electronic device.