A CNN-based multi-level face alignment approach for mitigating demographic bias in clinical populations

The investigation of demographic bias in facial analysis applications is a topic of growing interest with achievements in face recognition and gender classification. State-of-the-art convolutional neural networks (CNN) and traditional machine learning models for locating facial landmarks have reached overall performance levels close to human annotation. However, recent studies demonstrated that these models presented performance gaps when applied to different populations, characterizing bias led by demographic features. Nevertheless, few studies have addressed this problem in face alignment and facial landmarks localization methods. In this work, we propose a multi-level face alignment approach settled on CNN models to reduce performance gaps among different populations. We created facial subunit CNN models tied to a facial subunit detector at a higher level. The proposal seeks to improve bad results caused by facial impairment, guided by the following assumptions: facial unit landmarks localization does not require global texture, and combining different facial unit models can improve the final model’s variability. We applied the models in a balanced dataset mixing healthy controls and individuals with neurological disorders: amyotrophic lateral sclerosis and post-stroke. With fewer samples for training, our approach significantly reduced face alignment performance differences among those groups as compared to a state-of-the-art CNN model.