Designing an Enhanced MobileNet Integrated Architecture with Optimization for Accurate Detection of ASD using Extracted Facial Features

In the research, a new method is proposed to ensure the robust detection and extraction of facial features relevant for differentiating individuals with autism spectrum disorder. A multilevel cascaded Convolutional Neural Network (MCNN) architecture is designed for high-precision face localization in images, which can later provide a basis for subsequent fine-grained feature extraction from faces. Further, at the very initial stage of data preprocessing a very effective contrast adjustment (CA) technique is introduced that would efficiently enhance image quality and uniformity due to different lighting conditions. This improves, much better, the reliability of the facial features extracted and allows more accurate identification of discriminative emotional expression features for ASD. First, state-of-the-art high-precision facial landmark detection (FLD) techniques pinpoint key facial structures, thus helping extract relevant features for ASD classification. To further optimize feature selection, a Sand Cat Swarm Optimization (SCSO) method is introduced that identifies and maximizes subsets of facial features with maximal differences between ASD and non-ASD classes to boost classification accuracy. Finally, a customized hybrid Dense-MobileNet (DMN) architecture is provided that classifies ASD and non-ASD cases efficiently and accurately using the extracted facial features. This architecture synergistically puts together the advantages of dense layer to the MobileNet in order to guarantee high performance by achieving an accuracy of 99.22% in the identification of ASD-related features from the face.