Enhanced MobileNetV1 for Early Autism Detection: A Transfer Learning Approach Using Facial Imagery

The complex neurodevelopmental disorder known as autism, or autism spectrum disorder (ASD), is typified by repetitive behaviors, verbal and nonverbal communication difficulties, and ongoing challenges in social interaction. Since each person is affected differently and experiences symptoms that range from mild to severe, prompt and precise diagnosis is crucial for effective care. Automated solutions for ASD identification have emerged as a promising research area due to advancements in computer vision and machine learning. These solutions aim to bridge gaps in traditional diagnostic methods by offering faster and more accessible alternatives. They can also help minimize human error and variability in assessments, ensuring more consistent results. For individuals with ASD, early detection via automated methods can greatly enhance treatment outcomes and quality of life. Image-based analysis provides a scalable, objective substitute for time-consuming and biased traditional procedures such as behavioral evaluations and clinical observations. This paper presents a model that can detect autism using image processing, with MobileNetV1 employed as the feature extractor. The dataset used in this study is collected from the Kaggle website and contains a total of 2,940 images of autistic and non-autistic individuals. Advanced preprocessing techniques and classifiers are applied to enhance model performance. In terms of evaluation, the model's performance is measured using accuracy, precision, recall, F1-score, and sensitivity. The proposed method achieves an accuracy of 88.33%, with a lower error rate of 0.1167, which is comparably better than those reported in other studies.