• To introduce with Grid-wise expression features for recognizing ASD children. • To introduce grid-wise attention in ASDnet for learning long range dependencies. • To improve the speed of ASDnet by pyramid features using dual-branch fusion models. • To select-optimal hyperparameters using the HGSO algorithm that improve accuracy. • To validate the effectiveness of the proposed model by collecting real time images. Autism spectrum disorder (ASD) is a neurodevelopmental disorder related to brain growth and subsequently disturbs the physical appearance of the face. Therefore, it is crucial to create a learning model that can assist in ASD detection using facial expressions. In this work, the Dual-branch CNN-based visual transformation model (Db-CNN-VTM) is introduced to identify children with autism. Initially, the images collected from the dataset are subjected to pre-processing for denoising. A geometric data augmentation (DA) is introduced to perform the pre-processing by adding more training data. After pre-processing, feature extraction is performed by the Convolutional Neural Network with a grid-wise attention mechanism (CNN-GAM) model. After the feature extraction, the feature fusion residual network model is established. Finally, the classification is done using a novel Db-CNN-VTM model, and then the hyperparameter is found using the Hunger Game Search Optimization (HGSO) method. The obtained outcome for the developed model is finally compared with the other approaches, such as ResNet, AlexNet, CNN, and DNN models. Then, the performance of the proposed method is based on an accuracy of 96.91%, precision 96.91%, recall 96.89%, F1-score 96.90%, specificity 96.89%, FPR 0.03%, kappa 93.80%, MCC 93.80% and MSE 0.03.
Alamgir et al. (Mon,) studied this question.