Early detection of Autism Spectrum Disorder (ASD) is essential for effective intervention, yet conventional diagnostic approaches rely heavily on behavioral assessments that are subjective and time-consuming. This study presents a novel hybrid deep learning (DL) framework that integrates Histogram of Oriented Gradients (HOG) with the ResNet50V2 architecture for automated ASD classification using facial images. The proposed method leverages HOG for extracting local structural and edge-based features, which are then fused with deep hierarchical representations learned by ResNet50V2, enabling enhanced discriminative capability. The model is implemented using transfer learning, where pre-trained weights of ResNet50V2 are fine-tuned on the ASD facial image dataset. The HOG features are combined with the deep feature maps to form a robust feature vector, which is subsequently passed through fully connected layers for classification. This hybridization improves feature generalization and reduces overfitting compared to conventional CNN-based models. Experimental evaluation on a publicly available dataset demonstrates that the proposed approach outperforms existing state-of-the-art (SOTA) methods. The model achieves an accuracy of 90.7%, precision of 91.5%, recall of 91.0%, and F1-score of 91.2%. The improved performance is attributed to effective multi-level feature extraction and optimized network training. The results indicate that the integration of handcrafted and deep features provides a reliable, non-invasive, and efficient framework for ASD classification, highlighting its potential for deployment in automated screening systems. Furthermore, the proposed framework has the potential to assist healthcare professionals in decision-making and early diagnosis.
S et al. (Fri,) studied this question.