Scholarlink

<p>A hierarchical multi-scale attention framework for Facial Emotion Recognition is presented, addressing challenges such as subtle expression variations, class imbalance, and real-world image degradation. The framework fuses ResNet50 and Xception architectures and applies CBAM hierarchically across multiple semantic depths to recalibrate spatial and channel-wise features. A hybrid dual-pooling strategy captures both global context and local salient cues, improving discrimination of nuanced emotions. Evaluated on the FER2013 dataset, the framework achieves 78.9% accuracy and a 77.2% F1-score, significantly outperforming individual backbone models. Ablation and statistical analyses confirm the effectiveness of heterogeneous fusion, hierarchical attention, and pooling strategies. Despite its advanced design, the model maintains real-time performance with moderate complexity, demonstrating suitability for practical, embedded FER applications.</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

Hierarchical Multi-Scale Attention Framework with Heterogeneous Deep Network Fusion for Robust Facial Emotion Recognition

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Sudhakar Hallur,
  • Anil Gavade

摘要

A hierarchical multi-scale attention framework for Facial Emotion Recognition is presented, addressing challenges such as subtle expression variations, class imbalance, and real-world image degradation. The framework fuses ResNet50 and Xception architectures and applies CBAM hierarchically across multiple semantic depths to recalibrate spatial and channel-wise features. A hybrid dual-pooling strategy captures both global context and local salient cues, improving discrimination of nuanced emotions. Evaluated on the FER2013 dataset, the framework achieves 78.9% accuracy and a 77.2% F1-score, significantly outperforming individual backbone models. Ablation and statistical analyses confirm the effectiveness of heterogeneous fusion, hierarchical attention, and pooling strategies. Despite its advanced design, the model maintains real-time performance with moderate complexity, demonstrating suitability for practical, embedded FER applications.