Scholarlink

<p>The proposed Intersective Attention U-Net is a modified U-Net architecture that incorporates the Contextual Attention Branch (CAB) and Pose Normalization Branch (PNB) to enhance Facial Expression Recognition (FER), effectively handling pose variations, occlusions and subtle emotional cues. The CAB employs multi-head attention to selectively focus on expressive facial regions while suppressing occluded or irrelevant areas. Simultaneously, the PNB incorporates an Adaptive Spatial Transformer Network (ASTN) to dynamically align facial features, ensuring robustness to pose variations. Unlike traditional FER methods that struggle with misalignment and occlusions, the proposed fusion mechanism enriches feature representation, enabling the model to capture fine-grained emotional expressions with higher precision. Extensive evaluations on CK+, RAF-DB, and UTKFace datasets demonstrate the superiority of our approach, achieving 99.26%, 99.34%, and 99.43% accuracy, respectively, surpassing state-of-the-art FER techniques. The proposed framework offers a robust and adaptive solution for real-world 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

Facial Pose Adaptation: Intersective Attention U-Net for Facial Expression Recognition

  • 期刊论文

  • 发表日期：

    2026年3月20日

查看全文
  • Surya Pratap Yadav,
  • Shailendra Kumar Shrivastava

摘要

The proposed Intersective Attention U-Net is a modified U-Net architecture that incorporates the Contextual Attention Branch (CAB) and Pose Normalization Branch (PNB) to enhance Facial Expression Recognition (FER), effectively handling pose variations, occlusions and subtle emotional cues. The CAB employs multi-head attention to selectively focus on expressive facial regions while suppressing occluded or irrelevant areas. Simultaneously, the PNB incorporates an Adaptive Spatial Transformer Network (ASTN) to dynamically align facial features, ensuring robustness to pose variations. Unlike traditional FER methods that struggle with misalignment and occlusions, the proposed fusion mechanism enriches feature representation, enabling the model to capture fine-grained emotional expressions with higher precision. Extensive evaluations on CK+, RAF-DB, and UTKFace datasets demonstrate the superiority of our approach, achieving 99.26%, 99.34%, and 99.43% accuracy, respectively, surpassing state-of-the-art FER techniques. The proposed framework offers a robust and adaptive solution for real-world FER applications.