Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that is defined by persistent social interaction and communication difficulties, accompanied by limited interests, repetitive behaviors, and sensitivity to sensory input. Individuals with ASD tend to exhibit distinctive patterns in emotion recognition, facial expression, and response. This paper describes a comparative assessment of two deep models used for facial emotion detection in autistic children: an improved DenseNet121 and an innovative Hybrid Attention-based Explainable Network (HAXN). With a difficult dataset of 833 facial images corresponding to six emotions and characterized by severe class imbalance, we tested the performance of the models. DenseNet121 produced a superior overall test accuracy (65.33%) over HAXN (61.33%). Yet, while failing in test accuracy, HAXN showed better generalization on the validation set (72.22%) and provided better interpretability through its inbuilt attention mechanisms. Both networks were robust on most “joy” class (F1-score > 0.86) but broke down on minority emotions, especially “fear,” where HAXN completely collapsed (0% F1-score). Our results indicate: (1) traditional convolutional networks such as DenseNet121 still perform well on unbalanced emotion databases; (2) attention-based networks such as HAXN offer important transparency for clinical understanding; and (3) class imbalance still represents a major hurdle in autism-specific emotion recognition tasks. This work advances assistive technologies by identifying the trade-offs between explainability and accuracy in models. It further stresses the imperative for enhanced algorithms in dealing with minority-class learning, crucial to obtaining fair performance on all emotional classes.