Luca Saladino successfully defended his Master’s thesis titled „Skin Tone Estimation for Performance Differential Analysis in Biometric Systems.“ This is a joint effort of h-da (Germany) and DTU (Denmark).
Abstract:
This work develops a regression model that predicts continuous CIE LAB skin-tone values directly from a face image. Training is supervised by ground-truth measurements from a calibrated contact colorimeter. A ViT-S/16 backbone pretrained on ImageNet-21k is fine-tuned on the ICAO-compliant DAST-EX-10 dataset using a two-phase training schedule. When these continuous values are mapped to discrete bins by L* and B* nearest-centroid assignment, the model reaches 98.9% accuracy within one bin on a 10-class scale.
For more informations about current fairness-related research, the interested reader is referred to: https://impact.h-da.de/forschung-fuer-faire-biometrie