Researchers from ADAPT Dublin City University and the University of Galway have introduced ‘FaceSaliencyAug: Mitigating Geographic, Gender and Stereotypical Biases via Saliency-Based Data Augmentation’, a new technique to reduce biases in computer vision models. Developed by PhD student Teerath Kumar (CRT-AI & ADAPT Research Centre, DCU), Dr Alessandra Mileo (INSIGHT & I-Form Research Centre), and Dr Malika Bendechache (ADAPT & Lero Research Centres, University of Galway), this method uses saliency maps to mask key facial regions, increasing dataset diversity and improving model fairness.
Experiments show that FaceSaliencyAug reduces gender bias across diverse professions, as measured by Image-Image Association Scores, and in increasing dataset diversity, as measured by Image Similarity Scores. It outperforms traditional models like Convolutional Neural Networks and Vision Transformers in both accuracy and bias reduction. The approach has potential applications in facial recognition, healthcare, and HR.
Read the full paper here.
