An Autism Detection Architecture with Fusion of Feature Extraction and Classification

Impaired social and communication skills are hallmarks of the developmental disorder autism. Early detection of autism utilizing facial features and eye contact can aid in creating effective therapeutic strategies. This study evaluated a hybrid model using deep Convolutional Neural Network (CNN)-based transfer learning techniques to identify children with autism using Machine Learning and Deep Learning approaches. Variety of transfer learning models, such as ResNet101V2, Xception, DenseNet121, and others have been employed to extract features. The extracted features were then fed into classifiers like SVM, Logistic Regression, AdaBoost, KNN, and others. The ResNet101V2, combined with SVM and Logistic Regression, provided the highest accuracy of 88.5% and 87.5% respectively. The result analysis indicates that our proposed model performed significantly well in classifying autism with multi-valued features. This study can be well utilized by the researchers or doctors to increase the rate of such applications in future development.