HarmonyVisage: Ethical Facial Emotion Dataset Using Advanced Generative AI

Bias is a critical concern faced in AI development. One of the major causes is the discriminated dataset. To promote AI fairness and parity, this research built a novel ethical facial emotion dataset HarmonyVisage using various Generative AI models free of individual’s privacy and data rights. The dataset is balanced across ethnicities, genders, and age ranges based on the distribution ratio published by the U.S. Census Bureau. To promote diversity, image representations of neurodivergent and neurotypical populations are also included. To validate the fairness of the dataset, a trained FER AI model was used to assess the classification accuracy on the different population subgroups. The largest accuracy gap across all subgroups is 12.2%. The p-value between the White and African American subgroups is 0.250716, being insignificant. The consistent accuracy proves HarmonyVisage being a novel ethical dataset. This study not only contributes to the facial emotion recognition domain, but also makes a positive impact on other research domains where diversified human facial images are required.