A Comparison of Neural Networks for PCB Component Segmentation

In recent years, printed circuit board (PCB) assurance has become increasingly important. One method of addressing this need is through extracting the bill of materials (BoM) from an optical image of the sample and comparing it to reference parts from the board's design. This requires precise knowledge about mounted component shapes to properly account for various BoM properties. Semantic image segmentation, also known as pixel-level image labeling, is ideal for this task and already widely applied in a multitude of applications (e.g. medical, aerospace, geospatial, etc.). However, optical PCB images demonstrate characteristics which make it difficult to apply solutions from these alternative domains. This work describes the challenging nature of accurate PCB image segmentation and why existing solutions are not well-suited to meet these needs. Several recent techniques leveraging neural networks, namely UNet, DilatedNet, DeepLab, LinkNet, and ICNet are explored in their capabilities toward this purpose. Relevant impacts from a hardware assurance perspective are also analyzed.