Adapting Vision Transformers for Cross-Product Defect Detection

Advanced defect detection solutions that can easily adapt to different products and defect types are of high value for modern manufacturing companies. A significant challenge in developing and deploying such AI models is ensuring they generalize efficiently across diverse visual domains. This challenge is driven by limited data availability of high quality and the substantial effort required for labeling such datasets. This paper explores the adaptation of a Vision Transformer (ViT), originally trained to identify aesthetic defects in battery modules, for application in moulded plastic parts. By using transfer learning and generative AI techniques, this study evaluates fine-tuning and synthetic data augmentation strategies. The proposed approaches are assessed for their potential to enhance model adaptability and reduce dependency on extensive labelled datasets. A case study involving a battery manufacturing company with real-world data serves as the basis for this evaluation. Our preliminary findings suggest promising directions for enhancing the flexibility and efficacy of AI-driven defect detection systems in diverse manufacturing environments.