IDEmotion-GAN: Synthetic Data Augmentation for Facial Emotion Recognition in Intellectually Disabled Children

Facial Emotion Recognition (FER) has emerged as a crucial element in the fields of assistive technologies, healthcare systems, and intelligent educational platforms. Current FER models, however, are largely based on data from NONID adults, and are ineffective for children with ID, because of their limited sample sizes, atypical emotional expressions and high class imbalance. In this study, the authors introduce their data augmentation methods, namely data augmentation framework IDEmotion-GAN, which generates realistic synthetic facial images to enhance emotion recognition performance on intellectually disabled children. The framework combines three specialized datasets derived from autistic and intellectually disabled children and uses a Conditional Generative Adversarial Network (cGAN) to provide more examples of under-represented emotion categories. The augmentation pipeline consists of multiple steps, including rotation, reflection, translation, cropping, and GAN-based synthesis, to create a more diverse dataset. The experimental analysis shows that the proposed augmentation process is very effective in boosting the class balance and recognition accuracy than the traditional training methods. Emotion-specific characteristics in the synthetic samples are preserved while also increasing the intra-class variability for more robust training of the models. The proposed framework can be used in future to solve data scarcity problems in special needs emotion recognition systems, and can be useful for developing emotion-aware learning environments, emotion-aware therapeutic monitoring, and inclusive healthcare environments.