Child facial expression analysis (FEA) is a challenging problem due to (1) a limited amount of labeled data for children and (2) developmental differences in facial appearance making larger datasets of adult facial expressions an inappropriate ground-truth for child FEA. Transfer learning and domain adaptation are techniques aimed at improving upon the generalizability of models to a target domain (child facial expressions) that differs from the source domain (adult facial expressions) on which a model is trained. This project combines techniques from domain adaptation and deep transfer learning to classify images of facial expressions posed by children into seven classes (‘anger’, ‘disgust’, ‘fear’, ‘happy’, ‘sad’, ‘surprise’, plus ‘neutral’) while training with small number samples per expression. This project shows promise for improved generalized transfer learning with small data.

click image to enlarge



click image to enlarge