Face biometrics for differentiating typical development and autism spectrum disorder: a methodology for collecting and evaluating a dataset

Abstract

Autism Spectrum Disorder (ASD) is a neuro-developmental disability marked by deficits in communicating and interacting with others. The standard protocol for diagnosis is based on fulfillment of a descriptive criteria, which does not establish precise measures and influence the late diagnosis. Thus, new diagnostic approaches should be explored in order to better standardise practices. The best case scenario would be to have a reliable automated system that indicates the diagnosis with an acceptable level of assurance. At the moment, there are no publicly available representative open-source datasets with the main aim of this diagnosis. This work proposes a new methodology for collecting a Face Biometrics dataset with the aim to investigate the differences in facial expressions of ASD and Typical Developmental (TD) people. Thus, a new dataset of facial images was collected from YouTube videos, and computer vision-based techniques were used to extract image frames and filter the dataset. We have also performed initial experiments using classical supervised learning models as well as ensembles and managed to archive promising results.