Autism Spectrum Disorder (ASD) screening commonly combines behavioural observation, standardised instruments, and clinical judgement. Several early-associated behaviours are visually observable, motivating computer-vision-based screening support. This paper fine-tunes a Vision Transformer (ViT-base/16) on 5,866 images from sixteen autism-associated behaviour classes and compares it with a ResNet-50 baseline under the same protocol. On 645 held-out test images, ViT-base/16 achieves 89.92% accuracy, 90.91% precision, 89.92% recall, and 89.87% F1, compared with 63.26% accuracy for ResNet-50. Per-class results show strong performance for most behaviours, while Weird_Expression is the weakest class at 61.8% F1. A FastAPI backend and React frontend support uploaded images and OpenCV webcam inference, with a 50% confidence gate for uncertain predictions. Inference is performed on-device. The system is presented as a screening aid rather than a replacement for clinical diagnosis.

Autism Spectrum Disorder, Vision Transformer, transfer learning, behaviour recognition, image classification, real-time inference, OpenCV, confidence estimation