Leveraging Optimal Methods for Diverse Skin Types Classification in Images for Reduced Bias

This paper addresses the bias in automatic skin tone classification, highlighting how technologies can mirror societal issues such as inadequate representation of diverse skin tones. The objective is to evaluate methods that ensure balanced performance across skin types despite a scarcity of annotated databases with ethnic-racial details. Using the Fitzpatrick Skin Type classification system, we assess various skin color labeling algorithms, selecting optimal approaches for each type based on f-score. Results show that a single method biases results towards certain skin tones, whereas combining methods enhances accuracy across diverse types. We leverage the LFW and Fitzpatrick17k databases and apply image processing techniques like gamma transformation, CLAHE, histogram equalization, and non-linear order statistics filters. By tailoring processes to specific skin tone ranges, our auto-labeling approach better mirrors manual labeling, aiming for more equitable technology.