Emotion-based Autism Spectrum Disorder Detection by Leveraging Transfer Learning and Machine Learning Algorithms

Abstract

Autism Spectrum Disorder (ASD) presents as a neurodevelopmental condition impacting social interaction, communication, and behavior, underscoring the imperative of early detection and intervention to enhance outcomes. This paper introduces a novel approach to ASD detection utilizing facial features extracted from the Autistic Children Facial Dataset. Leveraging transfer learning models, including VGG16, ResNet, and Inception, high-level features are extracted from facial images. Additionally, fine-grained details are captured through the utilization of handcrafted image features such as Histogram of Oriented Gradients, Local Binary Patterns, Scale-Invariant Feature Transform, PHASH descriptors. Integration of these features yields three distinct feature vectors, combining image features with VGG16, ResNet, and Inception features. Subsequently, multiple machine learning classifiers, including Random Forest, KNN, Decision Tree, SVM, and Logistic Regression, are employed for ASD classification. Through rigorous experimentation and evaluation, the performance of these classifiers across three datasets is compared to identify the optimal approach for ASD detection. By evaluating multiple classifiers and feature combinations, this work offers insights into the most effective approaches for ASD detection.