Printed Circuit Boards at Perspectives: Towards Understanding Useful Data

Printed Circuit Boards (PCBs) are often the subject of computer vision tasks, within the realms of object detection and classification. This paper provides a new PCB dataset (PCB-P) that has a greater diversity in the range of perspectives and rotations that the images are captured at, as well as a diverse selection of PCBs. These specifications allow for data quality to be identified based on the configuration of the captured image. This dataset is tested on a small range of popular image classification architectures, finding Inception V3 to be the best-performing. Additionally, configurations of reducing various perspectives and rotations are tested. This work finds there to be diminishing returns from increasing training image quantity.