Résumé:
The study of autism spectrum disorder (ASD) is one of the most interesting research topics in
recent years because it requires a very rapid and accurate diagnosis. Specialists have relied on
children's ability to write, draw, and color, in addition to collecting special pictures of their
faces, which has provided many advantages in diagnosis. For example, children with autism
may have specific patterns in their drawings and writings, which can indicate certain aspects
of their condition. Analyzing these graphic elements can reveal fine motor skill difficulties or
repetitive tendencies typical of ASD. Moreover, the facial expressions of children can offer
clues about emotional recognition and social responsiveness, two areas often affected by
autism.
There are also several other methods of diagnosis, such as studying and analyzing behavior
via video, and relying on X-rays of the brain. Behavioral observation can help identify
specific behavioral patterns and communication difficulties. Additionally, brain imaging can
show abnormalities in brain structure and connectivity, providing biological evidence of
autism. For this reason, the diagnostic process remains somewhat complex and difficult. This
complexity has encouraged the use of modern artificial intelligence techniques, such as deep
learning, to diagnose autism spectrum disorders using coloring, handwriting, and drawing
photos and facial images. Artificial intelligence can analyze large amounts of data and
identify subtle patterns that humans might miss. To this end, artificial intelligence scientists
and researchers have begun to engage in building systems based on deep learning and
machine learning. These systems can process images and videos in a sophisticated manner,
learn to recognize the features associated with ASD, and improve the accuracy and speed of
diagnoses.
In this project, we created three educational models for autism spectrum classification. The
first model is an updated CNN model called (Aut-Net) that achieved an accuracy rate of 98%
on the coloring dataset. The second model, VGG 16, achieved 98% on the drawing dataset.
The third model, Res-Net, is related to the face recognition dataset and achieved 77%. These
models and datasets were chosen based on the results of experimenting with different models
and selecting the best one, in addition to the available datasets. This rigorous approach
ensures that the chosen models are the most effective for the given tasks, enhancing the
reliability and accuracy of the diagnostic process.
