Assessment of skin type measurement in image datasets

Skin type measurement is crucial for assessing whether image datasets used to train deep learning models, which lack skin type information, are sufficiently diverse. Earlier studies show that skin type measurement methods face challenges in accurately determining the true skin type across all skin tones and various lighting conditions and conclude that skin type measurement is a non-trivial task. As the most widely used method of skin type measurement, individual typology angle (ITA) is investigated here. The Fitzpatrick scale, which categorizes skin based on its color and reaction to ultraviolet light, was used as a ground truth for the performance assessment of ITA. This was done using the dermatologist-labelled data in the PAD-UFES-20 dataset which was modified to create a simplified skin patch dataset. The assessment results show a significant bias toward lighter skin types, a high misclassification rate for adjacent light skin types, and very few correct classifications for darker skin types. The main contributions of this work are: Establishment of a human skin patch dataset with Fitzpatrick skin labels as a benchmark for skin classification; Assessment of ITA performance across all six Fitzpatrick skin types; Initial testing of ITA as a skin type classification method.