Abstract:
Autism spectrum disorder ASD has remained a popular research area for the last
two decades. Autism is a severe spectrum disorder that affects communication,
intellectual and social skills. Each patient exhibits the same facial abnormalities, allowing
experts to diagnose the condition with just an image. When it comes to ASD
diagnosis from a facial features analysis perspective, key challenges include identifying
physical facial features, analysing data to try to correlate these observations
with ASD occurrence and classifying children’s images into autistic and non-autistic.
The focus of our thesis lies in the detection of ASD in children by facial morphology
without any medical tests and psychological therapy using only deep learning
techniques. We used the Autism Children dataset, which includes autistic and nonautistic
children facial images. We applied transfer learning approaches on various
pre-trained convolutional neural networks to extract features from autistic and nonautistic
children’s faces and classify them. With two distinct transfer learning based
approaches, we were able to determine the best feature extraction and classification
models for autistic and non-autistic children’s face images. MobileNet was employed
as a feature extractor and these features were then passed to the SVM classifier for
classification, which produced the best results and VGG-16 was fine-tuned with
three new dense layers. Both strategies outperformed existing techniques in terms
of accuracy. Our proposed method for the first screening of ASD patients saves both
money and valuable time.
