An Automated Optical Inspection System for PIP Solder Joint Classification Using Convolutional Neural Networks

Schmidt K, Rauchensteiner D, Voigt C, Thielen N, Bönig J, Beitinger G, Franke J (2021)

Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2021

Original Authors: K. Schmidt, D. Rauchensteiner, C. Voigt, N. Thielen, J. Bonig, G. Beitinger, J. Franke

Series: Electronic Components and Technology Conference

Book Volume: 71

Pages Range: 2205 - 2210

Conference Proceedings Title: IEEE

Event location: San Diego, CA US

ISBN: 978-1-6654-4097-4


DOI: 10.1109/ECTC32696.2021.00346


In the fields of electronics manufacturing, the application of through-hole devices is still required, as heat dissipation and high current carrying capacity plays an important role. To ensure the highest quality standards, these electronics production processes take a multitude of inspection processes into account. For the detection of error patterns regarding the quality of the solder connections, usually, high-end inspection machines are utilized in the industrial application. The Automated Optical Inspection is a commonly conducted process, using visible light and rule-based inspection routines, setup by process experts for the evaluation of the Region of Interest. The high overhead of creating and maintaining product-specific checking routines and machine acquisition leads to increased costs and severe dependency on expert know-how. A flexible inspection algorithm, implemented into low-cost equipment for image generation is expected to reduce acquisition and optimization costs, and lower dependency on expert knowledge and high-end machinery. In this contribution, we present a novel framework for the automatic, near real-time solder joint classification based on Convolutional Neural Networks, flexibly detecting, and classifying solder connections. We utilize existing Deep Learning architectures for detection and classification. The localization model utilizes a YOLO-architecture (you-only-look-once), learning feature inputs based on a supervised learning approach. Pseudo-labeling is carried out automatically by an anomaly detection model. The image generation is executed by an industrial low-cost camera and an industrial rack-PC. The developed prototype is integrated into the existing production infrastructure. The results indicate a satisfactory detection and classification of the investigated solder connections with the proposed system. Hence, this system represent an alternative to commercially available high-end inspection systems being used for an inline control of Pin-in-Paste and through-hole device solder connections.

Authors with CRIS profile

Involved external institutions

How to cite


Schmidt, K., Rauchensteiner, D., Voigt, C., Thielen, N., Bönig, J., Beitinger, G., & Franke, J. (2021). An Automated Optical Inspection System for PIP Solder Joint Classification Using Convolutional Neural Networks. In IEEE (Eds.), IEEE (pp. 2205 - 2210). San Diego, CA, US.


Schmidt, Konstantin, et al. "An Automated Optical Inspection System for PIP Solder Joint Classification Using Convolutional Neural Networks." Proceedings of the 2021 IEEE 71st Electronic Components and Technology Conference (ECTC), San Diego, CA Ed. IEEE, 2021. 2205 - 2210.

BibTeX: Download