Bias of facial analysis at NIR spectrum
This paper investigates the effectiveness of using NIR images to mitigate bias in facial recognition and gender classification systems. Two datasets, CASIA-Africa and Notre Dame (ND-VIL) were combined to create a dataset of NIR images that were balanced by race and gender. ResNet-50, SEResNet, LightCNN, and DenseNet-121 were trained on this dataset for facial recognition and gender classification. LightCNN was found to be the most effective model for performing both tasks. An analysis of the genuine and imposter distributions and the FMR-FNMR curves showed that white subjects performed better than black subjects and that male subjects performed slightly better than female subjects, but not in all instances. In gender classification, the best models showed to be slightly biased towards males. NIR images appear to show promise at helping lessen the degree of bias in these systems, and future work with a more consistent quality of images for black and white subjects could lead to a further reduction in bias.