| dc.description.abstract | Smart manufacturing refers to the use of digitalization for improving and automating
manufacturing processes. One use case is artificial intelligence (AI) used in quality
control, which can reduce production costs and heavy labor. Training AI models
requires large amounts of annotated data, which can be costly to obtain. This
study aims to examine whether generative adversarial networks (GANs) can be
used for improving an image classification model, which is commonly used in smart
manufacturing. The data used in the study consists of cropped photographs of
characters from serial numbers on automotive engine parts on a production line,
which can be used to link the parts to certificates used for quality control. The
GANs are trained on the real images. The generated images are then sampled to
a mixed set of synthetic and real images, on which a convolutional neural network
(CNN) is trained. In this study, we sample two small subsets of the total dataset,
and investigate how the size of the dataset affects the performance. Further, we also
show that this data can be well-represented by a low-dimensional subspace. This
property is used for developing specific methods for sampling synthetic data. The
study finds that using GANs for augmenting datasets can increase the performance
of the CNN significantly, even when the original dataset is small. Using sampling
methods based on subspaces is shown to have a positive effect when the number of
added synthetic samples is low, but random sampling yields a higher performance
otherwise. | en_US |
