Verifying the Applicability of Synthetic Image Generation for Object Detection in Industrial Quality Inspection

Abstract

Sparse and imbalanced data is a common challenge that practitioners must overcome when implementing industrial ML applications. This challenge concerns deep learning-based quality inspection systems in particular, as they often are obligated to adhere to high constraints in terms of reliability and performance. As deep learning quality inspection systems are usually implemented in a supervised manner, they additionally require balanced datasets that may be difficult or costly to obtain in production environments. However, new approaches using Generative Adversarial Networks for synthetic image generation promise a remedy by increasing the data amount of sparse classes, such as faults or defects. This paper presents an experimental use case where we employ a state-of-the-art image generator model of StyleGAN2 to a quality inspection application in laser beam welding to increase the number of defect images for training an object detector. We evaluate the generated images and their influence on the object detector’s performance using several training configurations. Our results reveal that with the limited amount of data, we are able to generate synthetic images that look promising at first glance. However, in the evaluation based on the object detector, we find that introducing synthetic images had an adverse effect on detection performance and robustness of the system. Further research is required to generate defect images from sparse datasets that can improve the performance of object detection systems in quality inspection.