AutNet: A personalized AI framework for emotion detection in children with autism using few-shot learning

Autism spectrum disorder (ASD) is a neuro-developmental condition hallmarked by impairments in an individual's social interaction, communication, cognitive, and motor skills. This hinders the ability of an individual to assimilate the facial expressions of their peers and express their own effectively. Generalized facial emotion recognition models developed for neurotypical children perform sub-optimally and cannot be generalized to autistic children due to this aberrant expression of facial emotion. Additionally, a severe lack of adequate training datasets for facial emotion expressions in ASD often impedes work in this domain. Despite the promising results shown by the multimodal emotion recognition systems, their reliance on multiple sensing modalities in sync is restrictive for use in practical assistive technology applications. Besides, the availability of annotated datasets of facial emotions in ASD is another issue in building a personalized recognition model. In order to overcome these limitations, this paper presents a novel framework called AutNet, which is based on deep embedding architectures for personalized facial emotion recognition in children with ASD. Rather than following the traditional learning approach using classification methods, AutNet uses deep metric learning and few-shot adaptation to learn discriminative embeddings from a limited number of training samples with high intra-class variability and inter-subject differences specific to facial emotion in ASD. In addition, to enable personalized learning process, we build a facial emotion image dataset using the publicly available SSBD video database of children with stimming behavior. The experimental results show that the proposed architecture obtains an average recognition accuracy of 81.3%, exceeding that of the traditional facial emotion recognition models in limited data scenarios. As far as we know, our work presents for the first time an embedding-based architecture for personalized facial emotion recognition in children with ASD.