Scholarlink

<p>The task of emotion recognition in autism-related affective computing applications is complicated task because of unusual facial expressions and extreme class imbalance in existing datasets. Minority emotions like fear and anger have been underrepresented and this has resulted in low classification performance. This paper aims to suggest a generative augmentation model based on the conditional Generative Adversarial Network (cGAN) on Facial Emotion Recognition Autistic Children (FERAC) dataset to generate more samples of fear and anger classes. Fréchet Inception Distance (FID) and Structural Similarity Index (SSIM) are used to validate the generated images so as to maintain quality and diversity. We trained ResNet152V2 convolutional neural network (CNN) and Swin Transformer models on both actual and GAN-Augmented datasets. The experimental findings indicate that F1-scores of fear improved by 20% and anger by +12% on GAN-Augmented dataset, and the overall accuracy increase by 4%. The analysis of confusion matrices proves the decrease in the misclassification of emotions of minorities. This work identifies GAN-based augmentation as a promising approach to both class imbalance and enhancement of emotion recognition systems that would be useful in autism-associated interventions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Emotion Recognition in Autistic Children through GAN-Based Synthetic Data Augmentation

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Hafiz Arslan Ramzan,
  • Wakeel Ahmad,
  • Syed M. Adnan

摘要

The task of emotion recognition in autism-related affective computing applications is complicated task because of unusual facial expressions and extreme class imbalance in existing datasets. Minority emotions like fear and anger have been underrepresented and this has resulted in low classification performance. This paper aims to suggest a generative augmentation model based on the conditional Generative Adversarial Network (cGAN) on Facial Emotion Recognition Autistic Children (FERAC) dataset to generate more samples of fear and anger classes. Fréchet Inception Distance (FID) and Structural Similarity Index (SSIM) are used to validate the generated images so as to maintain quality and diversity. We trained ResNet152V2 convolutional neural network (CNN) and Swin Transformer models on both actual and GAN-Augmented datasets. The experimental findings indicate that F1-scores of fear improved by 20% and anger by +12% on GAN-Augmented dataset, and the overall accuracy increase by 4%. The analysis of confusion matrices proves the decrease in the misclassification of emotions of minorities. This work identifies GAN-based augmentation as a promising approach to both class imbalance and enhancement of emotion recognition systems that would be useful in autism-associated interventions.