Scholarlink

<p>In recent years, Image Emotion Recognition (IER) has attracted increasing attention due to people’s preference for expressing emotions through images on the Internet. To enhance the performance of IER, existing methods mainly focus on extracting more discriminative emotional features from images, and then directly use the extracted features for the fine-grained emotion classification, overlooking the fact that emotion perception may be a progressive evocation process, which can significantly improve the performance of IER. Therefore, we propose a Vision Transformer-based Dual-Polarity Memory Network (ViDMNet) that simulates the progressive emotion evocation process to enhance the performance of IER. Specifically, we first utilize the shallow features of Vision Transformer (ViT) for polarity prediction, simulating the polarity perception during progressive emotion evocation process. Then, we construct a Dual-Polarity Memory Network (DPMN) to enhance polarity emotion features as discriminative cues to better guide the fine-grained classification. Finally, we fuse the enhanced polarity emotion features with the deep features of ViT to perform fine-grained emotion classification, simulating the progressive emotion evocation process, thereby significantly improving the performance of IER. Extensive experiments conducted on two emotional datasets, FI and EmotionROI, empirically demonstrate the superior performance of our proposed method compared to most existing methods, with achieved accuracies of 75.35% and 66.84%, respectively. Visualizations further prove the interpretability and effectiveness of our method. The code will be available at <a href="https://github.com/779602199/ViDMNet">https://github.com/779602199/ViDMNet</a>.</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

ViDMNet: vision transformer-based dual-polarity memory network for image emotion recognition

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Rui Zhang,
  • Zhongcheng Liang,
  • Huihui Li,
  • Xiaoyong Liu,
  • Guihua Wen

摘要

In recent years, Image Emotion Recognition (IER) has attracted increasing attention due to people’s preference for expressing emotions through images on the Internet. To enhance the performance of IER, existing methods mainly focus on extracting more discriminative emotional features from images, and then directly use the extracted features for the fine-grained emotion classification, overlooking the fact that emotion perception may be a progressive evocation process, which can significantly improve the performance of IER. Therefore, we propose a Vision Transformer-based Dual-Polarity Memory Network (ViDMNet) that simulates the progressive emotion evocation process to enhance the performance of IER. Specifically, we first utilize the shallow features of Vision Transformer (ViT) for polarity prediction, simulating the polarity perception during progressive emotion evocation process. Then, we construct a Dual-Polarity Memory Network (DPMN) to enhance polarity emotion features as discriminative cues to better guide the fine-grained classification. Finally, we fuse the enhanced polarity emotion features with the deep features of ViT to perform fine-grained emotion classification, simulating the progressive emotion evocation process, thereby significantly improving the performance of IER. Extensive experiments conducted on two emotional datasets, FI and EmotionROI, empirically demonstrate the superior performance of our proposed method compared to most existing methods, with achieved accuracies of 75.35% and 66.84%, respectively. Visualizations further prove the interpretability and effectiveness of our method. The code will be available at https://github.com/779602199/ViDMNet.