Face autistic classification based on thermal using image ensemble learning of VGG-19, ResNet50v2, and EfficientNet

The subject of this paper is the detection of Autism Spectrum Disorder (ASD) traits using facial recognition based on thermal images. The goal of this study was to evaluate and compare the performance of various Convolutional Neural Network (CNN) architectures in classifying thermal facial images of children with ASD, thereby facilitating the early identification of autistic traits. The tasks addressed include preprocessing a dataset of thermal facial images to prepare them for model training; conducting classification using three CNN architectures VGG-19, ResNet50V2, and EfficientNet; and assessing their performance based on accuracy, precision, recall, and F1-score metrics. The methods employed involved training these CNN models on a balanced dataset of 4,120 thermal facial images and splitting them into training, validation, and test sets. Each model underwent extensive training to determine its ability to effectively classify autism and non-autism classes. The results revealed that ResNet50V2 achieved the highest accuracy of 98.82%, followed by VGG-19 and EfficientNet with accuracies of 96.47% and 96.07%, respectively. ResNet50V2 also demonstrated superior generalizability due to its lower validation loss and higher classification accuracy compared to other architectures. Conclusion. The scientific novelty lies in: 1) introducing thermal imaging as an effective tool for detecting ASD traits; 2) demonstrating the superior performance of ResNet50V2 in classifying thermal facial images with high accuracy and generalization; and 3) exploring EfficientNet for the first time in this domain, highlighting its potential for improving autism diagnostic systems. This study contributes to advancing noninvasive methods for ASD detection and paves the way for further applications of deep learning in clinical diagnostics.