A Convolutional Neural Network Approach To Identify Presence Of Autism In Facial Images

Abstract

ASD (Autism Spectrum Disorder) is characterized as a neurological developmental disorder by the reduced ability of the children which affects their social interaction, language (or) behavioral skills. This influences to engage in repetitive and stereotypic behaviors. The underlying causes of ASD are still not well understood and diagnosed, but it is increasing at an alarming rate and the number of children suffering from this disorder are also rapidly increasing. Facial appearance (or) expression is the most powerful and natural non-verbal emotional communication method. They are very important in our day to day conversation, next to the tone of voice. But the kids with Autism Disorder may have the disability to communicate properly. Several existing systems aids in knowing the severity of the disorder and the child’s ability to present the facial expressions, to the verbal prompt provided. However, these systems work on the children having ASD. But there is a gap in identifying the existence of autism in the children at early stages. The proposed model addresses this issue based on facial expressions. The model detects the facial landmarks from the input image and extracts the key ROI (Region of Interest). Detection of facial landmarks from the input image is a part of the shape prediction issue. Given an input image, a shape predictor tries to localize the key points of interest (x, y coordinates) along the shape. These key points (features) are given as input to the CNN (Convolutional Neural Network) model and are trained to detect the presence of Autism in Kids at early stages. For experimental purpose Autistic Children Dataset available in Kaggle is used, and the implemented model is validated using internal validation as well as external validation with an accuracy of 81% and 73% respectively.