TinyDINO VHaar based Bi-directional Chameleon optimized LightASDNet for autism spectrum disorder detection

Autism Spectrum Disorder (ASD) in toddlers is characterized by neurodevelopmental deficits while its early detection remains challenging due to the lack of specific biomarkers and developmental variability. Therefore, a TinyDINO VHaar based Bi-directional Factorization with Chameleon optimized LightASDNet (TiDI-ASDNet) is proposed. In this framework, input images are initially pre-processed using Optimized Hierarchical Guided Image Filter (OHGF) to denoise and outlier removal, while Self-Distillation with No Labels version 2-Network Vector of Locally Aggregated Descriptors (DINOv2-NetVLAD) extracts the global visual patterns. Simultaneously, the input texts are pre-processed using One-Hot SMOTE (OHS) for easier interpretation. Besides, the standard structures interrupt syntactic/structural parsing due to autistic language corpora which clusters n-gram pattern, weakening the pragmatic deficits. To address these issues a Hybrid Tiny Encoder based HaarNet (TE-HNet) that extracts discriminative questionnaire features and maintains the lexical stringency. Since, the manifold torsion occurs due to discrepancy in encoding levels from non-isomorphic latent structures with non-diffeomorphic mapping, Bi-directional Encoder based Cross Factorization (BiE-xF) is employed and it learns about the shared semantics of visual features and behavioral linguistic features, which reduces Heteroscedastic Ambiguity. Besides, the motor stimming behaviour produces recurring self-stimulatory motor patterns that confound temporal alignment in vision-based ASD models. Thus, Modified Chameleon optimized LightASDNet (MC-LAN) is presented for classifying ASD and non-ASD thereby mitigates Dyspraxic Gait Aberrations. Simulations revealed the robustness of the framework with 99.2% AUC and 98.2% accuracy.