Ridwan Bello

Racial Bias in Facial Expression Recognition Datasets

Objective

This project evaluates the racial biases present in popular facial expression recognition (FER) datasets and assesses how these biases impact model performance and fairness across demographic groups. The goal is to promote equitable computer vision systems through better data curation and training.

Role & Contributions

• Audited FER datasets for racial representation disparities.
• Trained baseline FER models and measured performance gaps across racial groups.
• Experimented with fairness-aware training strategies and dataset balancing techniques.
• Documented the extent of bias and proposed mitigation recommendations.

Technologies & Methodologies

• Python, TensorFlow, Keras, OpenCV
• Fairness Metrics: Statistical Parity, Equal Opportunity
• Datasets: FER-2013, CK+, Custom-labeled datasets

Outcomes & Impact

The study revealed significant performance gaps, particularly underrepresented groups facing higher error rates. Proposed strategies include dataset augmentation and integrating fairness constraints during training. Findings contribute to fairness awareness in facial recognition technologies.

Related Publications & Presentations

• Ridwan O. Bello, Joseph D. Akinyemi, Khadijat T. Ladoja, Oladeji P. Akomolafe, "Racial Bias in Facial Expression Recognition Datasets: Evaluating the Impact on Model Performance", International Journal of Engineering and Manufacturing (IJEM), Vol.15, No.1, pp. 1-10, 2025. DOI:10.5815/ijem.2025.01.01 View Paper