GraphPCB: Graph-encoded Printed Circuit Board Datasets for Component Classification with Graph Neural Networks

ICLR 2026 Conference Submission14785 Authors

Keywords: Graph neural networks, PCB images, node classification, IC identification

TL;DR: We propose novel GNN datasets for application in IC classification on PCBs.

Abstract: We present a graph-based framework for Printed Circuit Board (PCB) image analysis, targeting core hardware assurance tasks such as IC segmentation and component identification. PCB images differ fundamentally from natural images in texture simplicity, spatial sparsity, and lack of informative backgrounds, limiting the effectiveness of traditional vision models. We propose GraphPCB, a generic scheme that transforms PCB images into graph-structured data, where nodes represent localized component regions and edges encode spatial proximity. This representation enables the application of Graph Neural Networks (GNNs) to PCB understanding, offering robustness to geometric variations and background noise. We release two high-quality GraphPCB datasets and analyze their structural properties, including graph heterophily and domain-specific challenges. Extensive experiments with various GNN architectures provide benchmarks and insights, establishing GraphPCB as a new testbed for node classification in structured visual domains.

Primary Area: datasets and benchmarks

Submission Number: 14785