Unsupervised facial skin type classification using CNN embeddings and SOM self-organizing maps

Mendoza Torres J.D.; Talledo Vega E.S.; Huarote Zegarra R.E. · Proceedings of the LACCEI international Multi-conference for Engineering, Education and Technology (2025) · DOI: 10.18687/leird2025.1.1.269

Resumen

This paper describes the design, development and implementation of a mobile application capable of identifying facial skin type (normal, oily, dry) from an image provided by the user. The solution is based on a hybrid system that integrates a convolutional neural network (CNN), used as a feature extractor, and a self-organizing network (SOM), in charge of classifying latent vectors into clusters representative of the skin type. The application architecture combines local processing (image capture and selection, skin tone selection) with remote services hosted in Hugging Face Spaces, accessible through a REST API. The model achieved accuracy levels above 90 % in controlled tests. The mobile implementation in Android Studio with Kotlin allowed to achieve a friendly and functional interface, compatible with modern devices. This approach proves to be an efficient, accessible and scalable alternative for automated dermatological assessment. © LEIRD 2025.All rights reserved.

Instituciones: Universidad Nacional Tecnológica de Lima Sur