Scholarlink

<p>The study aims to identify and classify subtle changes in facial expressions within video sequences. The proposed methodology begins with a novel Emotion Recognition Siamese Neural Networks (ER-SNN) algorithm to extract pertinent key-frames using L2-distance metrics to generate embedding vectors from identical CNN subnetworks. A contrastive loss function further discriminates embeddings to determine pertinent key-frames. The selected frames of facial movements are captured using a non-linear fractional order variational (NFOV) algorithm, which combines Charbonnier norm and Marchaud fractional derivative for robust results. The NFOV algorithm’s restricted color maps are further augmented to expand the training dataset, so that overfitting may be mitigated. These augmented color maps are subsequently utilized as input to pre-trained deep learning architectures (VGG16, VGG19, ResNet50) for comprehensive feature extraction. Ensemble learning techniques are employed to merge the pre-trained models, so as to enhance the overall accuracy. Evaluations conducted on the CK+, JAFFE, and CAER-S datasets reveal that the proposed methodology yields substantial improvements in facial expression recognition accuracy compared with the existing approaches.</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

Emotion in motion: harmonizing deep embedding, optical flow and transfer learning for video emotion analysis

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Navneet Gupta,
  • R. Vishnu Priya,
  • Chandan Kumar Verma

摘要

The study aims to identify and classify subtle changes in facial expressions within video sequences. The proposed methodology begins with a novel Emotion Recognition Siamese Neural Networks (ER-SNN) algorithm to extract pertinent key-frames using L2-distance metrics to generate embedding vectors from identical CNN subnetworks. A contrastive loss function further discriminates embeddings to determine pertinent key-frames. The selected frames of facial movements are captured using a non-linear fractional order variational (NFOV) algorithm, which combines Charbonnier norm and Marchaud fractional derivative for robust results. The NFOV algorithm’s restricted color maps are further augmented to expand the training dataset, so that overfitting may be mitigated. These augmented color maps are subsequently utilized as input to pre-trained deep learning architectures (VGG16, VGG19, ResNet50) for comprehensive feature extraction. Ensemble learning techniques are employed to merge the pre-trained models, so as to enhance the overall accuracy. Evaluations conducted on the CK+, JAFFE, and CAER-S datasets reveal that the proposed methodology yields substantial improvements in facial expression recognition accuracy compared with the existing approaches.