Análise de desigualdade de gênero em bases de dados de reconhecimento facial

Abstract

This paper addresses a critical issue related to gender inequality in datasets used for Facial Recognition (FR). While FR is a valuable tool, many of the datasets used in its training are disproportionately composed of images of male individuals, resulting in algorithms biased towards recognizing male characteristics. Due to the lack of a representative distribution in the data, these algorithms often experience a decrease in accuracy when applied to female faces. Therefore, the main objective of this research is to highlight how an unbalanced dataset can impair the performance of FR algorithms. To address this issue, we utilize pre-trained deep neural networks applying transfer learning techniques to improve gender representation and strive for an equalized algorithm. In this study, the UTKFace dataset and the convolutional neural network VGGNet were used to achieve the results. It is observed that with sample equalization, it was possible to improve gender-specific rates, thereby achieving a more balanced model accuracy. Transfer learning was instrumental in achieving this outcome, allowing a model pre-trained on a large collection of generic images to be adapted and refined specifically for the gender classification task on the UTKFace dataset.