Scholarlink

<p>Facial expression recognition (FER) is important for deciphering human emotions, with scope ranging from human-computer interaction to healthcare and affective computing. This research offers a new methodology that combines high-level preprocessing, data augmentation, feature extraction, selection, and classification methods to improve multiclass emotion recognition from facial images. A FER dataset is used for data collection, with a varied range of facial expressions. To counterpose and lighting changes while maintaining expression-specific information, Adaptive Expression-Preserving Normalization (AEP-Norm) is utilized for preprocessing. For data augmentation, Expression-Preserving Generative Augmentation (EPGA) is proposed, using a GAN-based pipeline to create realistic facial expression variations without distorting underlying emotional characteristics. Expression-Infused Graph Convolutional Network (EI-GCN) extracts robust and discriminative features from EPGA-augmented data for effective representation learning. Moreover, the Adaptive Graph-Attentive Feature Selection (AGAFS) approach uses mutual information to find the most expressive features for emotion classification. Lastly, the Graph-Embedded Transformer Classifier (GETC) is introduced, which integrates graph-structured feature representation with transformer-based self-attention to capture intricate relationships in the feature space. Comprehensive experiments and hyperparameter fine-tuning confirm the superiority of the proposed methodology in terms of accuracy, reaching a state-of-the-art 98.1%, outperforming current FER solutions. The outcomes underscore the importance of expression-preserving augmentation, graph-based feature extraction, and transformer-based classification in enhancing emotion recognition performance. This research provides a contribution towards developing FER through the establishment of an efficient framework with high accuracy in managing varied facial expressions and suitability for real-world implementation in emotion-aware systems.</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

Adaptive Expression-Preserving Normalization and Augmented Graph-Based Feature Learning for Facial Expression Recognition

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Nirmal Varghese Babu,
  • E. Grace Mary Kanaga

摘要

Facial expression recognition (FER) is important for deciphering human emotions, with scope ranging from human-computer interaction to healthcare and affective computing. This research offers a new methodology that combines high-level preprocessing, data augmentation, feature extraction, selection, and classification methods to improve multiclass emotion recognition from facial images. A FER dataset is used for data collection, with a varied range of facial expressions. To counterpose and lighting changes while maintaining expression-specific information, Adaptive Expression-Preserving Normalization (AEP-Norm) is utilized for preprocessing. For data augmentation, Expression-Preserving Generative Augmentation (EPGA) is proposed, using a GAN-based pipeline to create realistic facial expression variations without distorting underlying emotional characteristics. Expression-Infused Graph Convolutional Network (EI-GCN) extracts robust and discriminative features from EPGA-augmented data for effective representation learning. Moreover, the Adaptive Graph-Attentive Feature Selection (AGAFS) approach uses mutual information to find the most expressive features for emotion classification. Lastly, the Graph-Embedded Transformer Classifier (GETC) is introduced, which integrates graph-structured feature representation with transformer-based self-attention to capture intricate relationships in the feature space. Comprehensive experiments and hyperparameter fine-tuning confirm the superiority of the proposed methodology in terms of accuracy, reaching a state-of-the-art 98.1%, outperforming current FER solutions. The outcomes underscore the importance of expression-preserving augmentation, graph-based feature extraction, and transformer-based classification in enhancing emotion recognition performance. This research provides a contribution towards developing FER through the establishment of an efficient framework with high accuracy in managing varied facial expressions and suitability for real-world implementation in emotion-aware systems.