ASD is a neurodevelopmental condition with significant challenges in social communication, and early intervention is crucial. In developing countries like Peru, more than 90% of ASD cases remain undiagnosed, highlighting the need for alternative screening tools. The proposed system uses artificial intelligence (AI), machine learning, and computer vision to analyze children’s emotional expressions, which can serve as early indicators of ASD. The goal is to provide an accessible and low-cost tool rather than replace professional diagnoses. The methodology involved using the “Autism_Image_Data” Kaggle dataset, which contains facial images of children with and without ASD. Preprocessing steps included resizing, image conversion to RGB, normalization, and data augmentation. Six deep learning models: SimpleCNN, ResNet50, EfficientNetB0, DenseNet121, Swin Transformer, and Vision Transformer (ViT) were implemented and evaluated under the metrics: accuracy, precision, recall, and F1-score. Initial results without hyperparameter tuning showed that ResNet50 had the best performance, with an accuracy of 91%. After tuning, EfficientNetB0 achieved the highest performance, with an accuracy of 94%, precision of 94%, recall of 93%, and an F1-score of 93%. The area under the curve (AUC) of the tuned model also improved to 0.921. These findings suggest that EfficientNetB0 is both accurate and computationally efficient, making it a promising model for real-world ASD early detection applications.