Improving Facial Expression Recognition through Data Preparation & Merging

Human emotions present a major challenge for artificial intelligence. Automated emotion recognition based on facial expressions is important to robotics, medicine, psychology, education, security, arts, entertainment and more. Deep learning promising capturing complex emotional features. However, there no training dataset that large representative of the full diversity in all populations contexts. Current datasets are incomplete, biased, unbalanced, error-prone have different properties. Models learn these limitations become dependent specific datasets, hindering their ability generalize new data or real-world scenarios. Our work addresses difficulties provides following contributions improve recognition: 1) methodology merging disparate in-the-wild increases number images enriches people, gestures, attributes resolution, color, background, lighting image format; 2) balanced, unbiased, well-labeled evaluator dataset, built with gender, age, ethnicity predictor successful Stable Diffusion model. Single- cross-dataset experimentation show our method generalization FER2013, NHFI AffectNet by 13.93%, 24.17% 7.45%, respectively; 3) we propose first largest which can complement real tasks related expression.