Deep Learning Based UV Facial Imaging Generation

Abstract

Skin health has become a topic of interest in the recent years. To ensure a better diagnosis and treatment, the analysis of high-quality skin databases is crucial. In this regard, UV imaging is a valuable tool in detecting melanoma and other skin conditions. However, UV images present some challenges both in availability and processing. For this reason, in this work we present UVnet, a method to generate optical-to-UV facial images based on autoencoder architectures. The proposed UVnet is validated across an extension of the Baependi Heart Study and other state of the art method. Our proposal successfully generates pseudo-UV samples with an average RMSE of 0.0040 and a structural similarity index against the actual samples of 0.2984. These results show that UVnet consistently achieves higher sample quality than existing methods and provides new capabilities regarding generation of large areas of the facial epidermis. This can be regarded as an initial effort to provide affordable access to high-quality skin databases.