Desenvolvendo um database equalizado para o reconhecimento de face

Abstract

The use of face recognition algorithms has been adopted extensively in several areas. However, a significant challenge that still needs to be addressed is the presence of discriminatory racial bias in these systems due to the set of images that make up their databases. It is important to emphasize that racial bias in image databases is due to several factors, including historical inequalities, inadequate representation and discriminatory trends in previous data collections. In this context, the present study proposed the use of balanced image databases, respecting racial diversity, aiming to increase the accuracy of facial recognition results and minimize discriminatory bias. In the work, 3 databases of public face images were used, namely CelebA with 202,599 face images, FEI Database with 2,304 images and UTKFace with about 20,000 images. Segmentation techniques with Thresholding were applied in order to separate pixels that correspond to skin color in an image in the three databases and thus classify the input images according to skin color. Resulting in this way in the new equalized database, developed with 6,000 images of people with white, yellow and black skin represented in an equal way, built with images of male and female faces of different ages.