Paper 2020/366

FICS-PCB: A Multi-Modal Image Dataset for Automated Printed Circuit Board Visual Inspection

Hangwei Lu and Dhwani Mehta and Olivia Paradis and Navid Asadizanjani and Mark Tehranipoor and Damon L. Woodard

Abstract

Over the years, the computer vision and machine learning disciplines have considerably advanced the field of automated visual inspection for Printed Circuit Board (PCB) assurance. However, in practice, the capabilities and limitations of these advancements remain unknown because there are few publicly accessible datasets for PCB visual inspection and even fewer that contain images that simulate realistic application scenarios. To address this need, we propose a publicly available dataset, FICS-PCB, to facilitate the development of robust methods for automated PCB visual inspection. The proposed dataset includes challenging cases from four variable aspects: PCB manufacturing, illumination, scale, and image sensor. The FICS-PCB dataset consists of 8,685 images of 31 PCB samples and contains 75,965 annotated components. This paper reviews the existing datasets and methodologies used for PCB visual inspection, discusses problem challenges, describes the proposed dataset, and presents baseline performances using feature-based and deep learning methods for automated PCB component visual inspection.