Análise comparativa de bases de dados etnicamente equalizadas e não equalizadas para reconhecimento de faces usando redes neurais convolucionais

Abstract

This study investigates algorithmic bias in facial recognition models, focusing on the impact of ethnic equalization in datasets on the performance of convolutional neural networks (CNNs). The research compares two datasets: FairFace, which has a balanced ethnic distribution and seven classes, and UTKFace, which is imbalanced across its five racial classes. By training the CNN and comparing training speed and accuracy, as well as testing metrics—precision, recall, F1-score for each category, and the standard deviation of these metrics—the results demonstrate that diversity and balanced racial classes are crucial factors in mitigating algorithmic bias in CNN-based facial recognition systems. The findings highlight that the training speed and equalization of outcomes improve with more representative and racially balanced datasets. This work underscores that using datasets that prioritize representativeness and racial balance in machine learning model training is essential for developing fair and equitable facial recognition systems, ultimately preventing the perpetuation of inequality.