This research aims to develop a general-purpose visual inspection AI that can handle diverse products and inspection items using a large-scale visual language model (LVLM). Conventional AI visual inspection requires collecting large amounts of training data for each product and defect type and building individual models, resulting in high implementation costs. This research realizes a technology that can handle new products and unknown defects by providing only a small number of good and defective product images and inspection instructions in natural language. Furthermore, it will develop functions that not only determine whether an item is good or bad, but also identify the defective area and explain the reasoning behind the judgment in natural language. This will enable easy implementation even for users without specialized knowledge and promote inspection automation in a wide range of manufacturing sites, including high-mix, low-volume production. In the future, the goal is to disseminate the technology domestically and internationally through commercialization.

Kunihito Kato
Professor, Gifu University