Responsible AI in Action: Enhancing Fairness in Computer Vision through Model Improvement Strategies

As facial recognition technology continues to advance, addressing issues of fairness and accuracy in image datasets becomes increasingly critical. This paper outlines a novel approach aimed at simultaneously improving the accuracy of facial recognition systems and promoting fairness within image datasets. Traditional approaches have often struggled to balance these objectives, with accuracy improvements potentially exacerbating biases. In our proposed methodology, we have proposed a single input multi-output model with non-linearity to extract the best feature map to classify age, gender and ethnicity to refine facial recognition models. Our experiments demonstrate that our approach enhances overall accuracy for gender and race and reduces mean absolute error (mae) for age. This research contributes to the ongoing discourse on fairness in AI by providing practical insights by improving the performance of facial recognition systems across diverse demographic groups. We have tested our model on UTKFace dataset and compared its performance with existing CNN models and show cased that to the best of our knowledge, it is achieving least mae for age than state-of-the-art models and almost comparable accuracy for gender and ethnicity without overfitting which is the case for many other existing CNN models.