Autism face detection is a crucial problem in today’s era. With the advancement in computer vision and pre-trained models, it is now easy to utilize less available images and with the help of weights of already trained model in the similar area. We compared various techniques of handcrafted feature extraction with pre-trained model and provide a comprehensive overview. Our fine-tuned VGG16 achieved 87% avg. accuracy on test dataset. EfficientNET-B0 and EfficientNET-B7 provides 79.3% avg. and 82.6% accuracy on test dataset. Furthermore, we tested DenseNet-121 and SE-ResNET-152 pre-trained models and after performing fine tuning we got 82.67% avg. and 85.3% accuracy on test dataset. Moreover, we used MobileNetv1 which achieved 87.0% avg. accuracy on test dataset. Furthermore, to increase the performance of our model we tried ensemble of three pre-trained model i.e. MobileNET, Xception and SE-ResNet, and got the best 89.19% accuracy.