Influência do pré-processamento de imagens no reconhecimento facial e seus vieses algoritmos

Abstract

The advancement of artificial intelligence technologies has driven the adoption of facial recognition systems in various social contexts, raising debates about their ethical and social impacts, particularly concerning algorithmic racism. This study aims to investigate the influence of image preprocessing on the performance of facial recognition models and its relation to the occurrence of algorithmic biases. As a theoretical foundation, the concept of algorithmic racism is adopted, understood as a digital update of structural racism in which historical inequalities are reproduced within technological processes that are often presented as neutral. In this sense, automated practices such as facial recognition can reinforce racial stereotypes and exacerbate social vulnerabilities. Another key concept is computer vision, a field of artificial intelligence dedicated to enabling machines to interpret visual data. Within this field, image processing—particularly adjustments in brightness and contrast—plays a crucial role in determining the accuracy and fairness of recognition models. For this purpose, 35,000 images were selected from the FairFace dataset, equally distributed across seven racial groups, allowing for controlled experiments without the introduction of initial imbalances. The methodology consisted of training convolutional neural network models based on the MobileNetV2 architecture, implemented using the TensorFlow library in Python. The dataset was subjected to two distinct preprocessing approaches: traditional histogram equalization and Contrast Limited Adaptive Histogram Equalization (CLAHE). In addition, data augmentation and regularization techniques were applied to enhance model robustness. The results show that CLAHE outperformed traditional histogram equalization, achieving 56.11% accuracy compared to 30.74% for the histogram-based model. Significant performance variation among racial groups was also observed, demonstrating that even in balanced datasets, preprocessing can introduce algorithmic bias. For instance, the black and east_asian classes achieved higher recognition rates with CLAHE (75.60% and 68.40%, respectively), while the latino_hispanic group showed considerably lower performance (24.40%). It is concluded that preprocessing techniques play a decisive role in facial recognition performance and may either mitigate or intensify racial biases in results. Thus, this study contributes to a critical understanding of preprocessing impacts on biometric systems and emphasizes the importance of responsible practices in the development and deployment of computer vision models aimed at reducing algorithmic racism.