Enhancing Autism Spectrum Classification using Super Resolution GAN and ViT based Embedding

Autism spectrum disorder(ASD) poses challenges in recognizing and socializing with others very challenging. Early detection is crucial for improved outcomes in individuals with ASD. In this research we present a novel approach for autism classification using facial images. Our approach utilizes the GAN based super-resolution techniques in conjunction with a Pretrained ResNet18 and vision transformer based embedding by combining their features. Our study focuses on enhancing image quality through super resolution, enabling finer details to be captured for mode accurate analysis. CNN based ResNet18 captures the local features from image while ViT, attention based network captures more global details. By combining the features of both networks our propose model performs robust and discriminative feature extraction. The integration of super-resolution techniques and fine-tuned deep learning models provides a comprehensive framework for early detection of autism in children. To correctly detect the autism disorder, more robust ResNet18 features and vision transform based embedding are extracted from enhanced image details. The enhanced image quality aids in capturing subtle facial expressions, critical for accurate classification. Our findings demonstrate the effectiveness of our proposed model in achieving notable improvements in accuracy and reliability with test accuracy of 91% and AUC of 96.42%. This research contributes to the growing body of knowledge in autism classification, highlighting the potential of advanced computer vision techniques and deep learning algorithms. By enabling early detection, out approach provides a positive impact on the lives of children with ASD, facilitating timely interventions and support for better outcomes and increased independence.