Comparative Study on Different Methods to Generate Synthetic Data for the Classification of THT Solder Joints

Automated Optical Inspection (AOI) is still one of the major tools for the quality control of solder joints; especially, if the requirements for the solder joints' mechanical and electrical properties are high. This is the case for products in the automotive sector, e.g. air bags or brake systems. Conventional test routines are time-consuming to define and not flexible regarding quality limits and lead to a high amount of manual reinspection. Thus, Artificial Intelligence offers the potential to address these disadvantages. Artificial Intelligence, on the other hand, suffers from highly imbalanced datasets resulting from the low amount of defects on solder joint level. One possibility to face this challenge is the utilization of synthetic data. This work compares different methods to generate data using both conventional data augmentation and deep generative models as well as virtual world rendering in order to enhance dataset quality in general. These are applied to an industrial dataset of THT solder joints. All tested approaches lead to an improvement in model quality. The best results are achieved with synthetic images from a Generative Adversarial Network, relatively increasing performance by around 6 % while minimizing error slip.