Towards PCB Netlist Extraction from Multimodal Imagery

Modern technology relies on the integrity of commercial-off-the-shelf (COTS) printed circuit boards (PCB), which often offer high performance at low cost compared to custom-designed alternatives. Due to the lack of visibility into the global supply chain, the manufacturing of these electronics can introduce deviations from the original design. One way to ensure that these devices functions as intended is to non-destructively decompose the manufactured assembly into a description of its design, often in the form of a netlist. However, automated decomposition of PCBs is non-trivial, as components, pins, and traces may be obscured, hindering the efficacy of the netlist extraction process. Therefore, we propose a surface imaging solution to extract the PCB netlist while overcoming these challenges. We formulate extraction as a segmentation problem and train a Feature Pyramid Network (FPN) to identify their locations on the board. Experimentation shows that the multimodal imaging improves the models' ability to correctly segment traces and pins. We also show that the choice of loss function is critical to successfully applying an FPN to this extraction problem.