Deep Transfer Learning-Based Emotion Recognition for Autistic Children

Sheila Agnes Vidot, J. Angelin Jeba, S. Rubin Bose, R. Regin, P. Paramasivan Department of Medical Consultants, IPS Health, Mahe, Seychelles. Department of Electronics and Communication Engineering, S.A. Engineering College, Chennai, Tamil Nadu, India. School of Computer Science and Engineering, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Research and Development, Thangavelu Engineering College, Chennai, Tamil Nadu, India.

Abstract: Autism spectrum disorder is a complex brain development disease that impacts how people view and interact with others. This hinders learning, communication, and conduct. Autism is called a "developmental disorder" because symptoms show in the first two years of life but can be diagnosed at any age. Common ASD symptoms include Poor social skills, repetitive conduct, and symptoms that impair school and work performance. Autism is called a "spectrum" disorder because its symptoms, needs, strengths, and problems vary. This study aims to rapidly assess patients' emotional states with ASD to provide individualised care and improve quality of life. However, due to minor differences in emotional expression, emotion detection in autism is difficult. A deep learning system for facial emotion identification, utilizing augmented facial image datasets, solves these problems. Initial multi-class classification using MobileNetV3Small classified all six emotions (Natural, Anger, Fear, Joy, Sadness, and Surprise). After unfreezing 46 trainable variables, the model's accuracy was 60.13% after 20 epochs with significant class imbalance. After this difficulty, researchers opted to binary-classify “joy” and “sadness” using EfficientNetV2Bo, the best-performing model in the initial set. The novel method achieved remarkable results, including 91.77% training accuracy and 95.00% validation accuracy. To enable high-accuracy multi-class classification across all emotion categories, future research will address strategies to address class imbalance.

---

Keywords: Autism Spectrum Disorder; Convolutional Neural Networks; Artificial Neural Network; Facial Expression Recognition; Typically Developing Children; Squeeze and Excitation; MobileNetV3Small.

---

Received on: 23/07/2025Revised on: 28/09/2025Accepted on: 01/12/2025Published on: 07/03/2026

---

---