Multi-class classification for image groups with ambiguous boundaries: Case study of skin wrinkle grading

Abstract

In the cosmetics industry, there is the demand for products with anti-wrinkle effects. Assessors such as physicians, evaluate efficacy by visual skin or images, and instrumental measurements, but there are debates about reliability. Therefore, previous studies have tried to develop models which objectively classify the wrinkle grades from images, but it hasn’t achieved sufficient performance. While the medical industry has shown the benefits of using natural language for image diagnosis supports by artificial intelligence. Accordingly, this study aims to develop a high-performance model that can classify to eight grades for skin wrinkle images with ambiguous boundaries and examine benefits of applying natural languages. We confirmed that adopting this technique contributes to improving the generalization and stabilization of models. Furthermore, implementing this approach enables constructing visual question answering models. In the future, we will examine the quality of language to further improve the performance of the skin wrinkle classification model.