Skin Tone Assessment Using Hyperspectral Reconstruction from RGB Image

Abstract

Fitzpatrick skin type (FST) classification, the most common skin tone measure, requires subjective assessment by dermatologists who ask the subjects about their ethnicity, skin response to sun exposure and medical history. FST may be inaccurate due to recall bias in self-reported answers and subjectivity of assessment. Although existing computer vision methods can objectively classify skin types using red green blue (RGB) images of skin, they are limited by spectral resolution of the RGB images and therefore inaccurate. Moreover, existing objective methods have little correlation with FST. In this work, we investigate computational methods of hyperspectral (HS) reconstruction from RGB images of skin. We further train novel skin type classification models based on reconstructed HS images of skin and evaluate them on a clinical dataset. Proposed models outperform RGB image based models such as individual typology angle significantly. This work illustrates that HS reconstruction of a skin image is much more useful than the corresponding RGB image in a wide range of skin related applications including cosmetics, dermatology and biometrics, while providing an inexpensive and easily accessible alternative to specialized HS imaging systems.