Enhancing the identification of autism spectrum disorder in facial expressions using DenseResNet-Based transfer learning approach

Autism Spectrum Disorder (ASD) is a neurological disability, which is characterized by cognition, behavioral challenges, social skills, and communication. Identifying children with ASD in the early stages helps improve their learning ability and limit the symptoms. Since there is no medical test to diagnose the disorder, it is very challenging. Many researchers diagnose ASD through facial images and the child's behavior. In this methodology, DenseResNet (DRN)-based transfer learning approaches are proposed to analyze the facial images through deep learning techniques. Generally, the human face reflects the human brain and when it is used as a biomarker, it becomes easy to diagnose ASD in its early stages. This strategy combines two methods: densely connected networks and residual networks, which are pre-trained models to detect facial images of autistic children. This model reaches 97.07% classification accuracy and uses its dataset. The deep learning model creates four dense blocks for extracting features. Moreover, it extracts the features from residual network layers and finally combines the features for classifying the images. This model is trained using around 2526 images and is tested using 200 images. Based on classification accuracy, the autism diagnosis system for children is effectively used by using only facial images.