A comparison of Vision Transformers and CNNs for the detection of aesthetic defects

The early detection of aesthetic failures in manufacturing is important for preserving customer satisfaction and business reputation, but also to avoid material wasted in producing defective products. This work discusses a vision-based system for the detection of aesthetic defects using Vision Transformers (ViT). Their performance is compared against a Convolutional Neural Network (CNN) model, assessing their suitability, in terms of accuracy and execution time, for proactive quality control in manufacturing. In both models, K-means clustering has been employed, for unsupervised image labelling through image segmentation. The two models have been evaluated with data from a battery manufacturing line.