Classificação de gênero com redes neurais convolucionais: uma análise comparativa aplicada a um dataset piauiense

Abstract

Classifying biological sex from facial images is a challenging task, often compromised by demographic biases that affect the accuracy of models in specific populations. Models trained with international datasets, for example, show reduced performance when ap- plied to populations with distinct facial characteristics, such as the Brazilian population. This monograph proposes a comparative analysis of five different Convolutional Neural Network (CNN) architectures for this task. To mitigate demographic biases, the method- ology was based on the transfer learning technique, applied to a state government database of 33,033 facial cutouts of individuals from Piauí, with the aim of determining the most effective architecture for sex classification. The results obtained highlighted the superiority of the ResNet50 model, which achieved the best performance indices, with values of 0.8957 (Accuracy), 0.7905 (Kappa), 0.8958 (Weighted F1-Score), and 0.9609 (AUC-ROC). However, the qualitative error analysis revealed that the models developed a strong correlation between the presence of cosmetic features and gen- der classification, indicating the persistence of biases related to the task itself and originating from the data. These results demonstrate the potential of the approach to create systems more adapted to local diversity and emphasize the importance of careful architecture selection and continuous data curation for the development of more accurate and equitable facial recognition technologies.