![Faculty of Engineering and Computing header](/sites/default/files/styles/max_1300x1300/public/scholarship_editor/2022-11/FEC_Faculty%20of%20Engineering%20%26%20Computing.png?itok=_kEIJkOo)
![DCU Campus](/sites/default/files/styles/user_defined/public/engineering_and_computing_editor/2025-02/dcu-glasnevin-22th-sept-2021-_8_0.jpg?itok=ZIqo_drr)
FaceSaliencyAug: A Novel Approach to Mitigate Biases in Computer Vision Models
This research study presents FaceSaliencyAug, a novel data augmentation technique designed to mitigate geographical, gender, and stereotypical biases in computer vision models. The method utilises saliency maps to identify and mask salient facial regions, enhancing dataset diversity and improving model fairness.
Experiments demonstrate the efficacy of FaceSaliencyAug in reducing gender bias across diverse professions, as measured by Image-Image Association Scores, and in increasing dataset diversity, as measured by Image Similarity Scores.
The investigation compares the approach's performance with Convolutional Neural Networks and Vision Transformers, revealing improvements in both accuracy and bias reduction. The study discusses real-world applications in facial recognition, healthcare, and human resources.
Read the full paper here.