Automated Structure Matching of Printed Circuit Boards

We address the challenge of matching photos of printed circuit boards (PCBs) that have the same layout. A simple approach to this problem is to use an off-the-shelf computer vision (CV) machine learning (ML) model to embed images into vectors and then calculate the similarity between those embeddings. However, this approach is prone to diversion by visual differences between photos that do not affect PCB layout. With a large, paired dataset of matching PCBs with visual differences, it would be possible to train a ML model to ignore these visual differences, but such a dataset does not exist and is expensive to create. We develop an approach that alleviates this problem by considering only component bounding boxes and classes, as outputted by an object detection ML model, forcing it to ignore irrelevant visual features. Specifically, we use a structure matching algorithm that compares the type, shape, and relative position of the components on the PCB. The algorithm uses the layout of the components as a proxy for the overall PCB layout. We demonstrate that our method can be used in conjunction with visual embedding similarity in order to take advantage of both visual similarities and structural similarities between PCB images. Our approach also has two unique capabilities. First, it finds the optimal alignment of the images such that matching components are overlaid. Second, in addition to photos, it can be directly applied to component data without the need for an object detector. This leaves the application of this algorithm open to manually inputted or modified component structures or schematics. We additionally open-source our evaluation and object detection code, model weights, and manually-gathered matched dataset used for evaluation at https://github.com/twosixlabs/pcb structure matching.