Scholarlink

<p>Facial Beauty Prediction (FBP) presents a challenging perceptual regression task requiring nuanced understanding of both fine-grained spatial features and high-level semantic concepts. Traditional deep learning models often excel in one area at the expense of the other. Here, we propose the Hybrid Vmamba-ViT architecture, which synergistically combines the strengths of two cutting-edge vision models: a Vmamba branch for fast and efficient long-range spatial feature extraction using State Space Models (SSMs), and a Vision Transformer (ViT) branch leveraging a pretrained backbone for deep semantic awareness through global self-attention. The complementary feature sets from these pathways are intelligently integrated using a multi-head cross-attention fusion mechanism, and the fused representation is passed to a regression head to predict a continuous beauty score. Evaluated on the SCUT-FBP5500 dataset using a rigorous 5-fold cross-validation protocol, our model achieves state-of-the-art performance with a Pearson Correlation (PC) of 0.9261. These results demonstrate the superior capability of our hybrid approach in capturing the complex perceptual cues essential for robust facial beauty assessment. The code is available at <a href="https://github.com/DjameleddineBoukhari/Hybrid-Vmamba-ViT-FBP">https://github.com/DjameleddineBoukhari/Hybrid-Vmamba-ViT-FBP</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

Enhancing Facial Beauty Prediction via a Dual-Pathway Hybrid Architecture Integrating Vmamba and ViT

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Djamel Eddine Boukhari,
  • Ali Chemsa

摘要

Facial Beauty Prediction (FBP) presents a challenging perceptual regression task requiring nuanced understanding of both fine-grained spatial features and high-level semantic concepts. Traditional deep learning models often excel in one area at the expense of the other. Here, we propose the Hybrid Vmamba-ViT architecture, which synergistically combines the strengths of two cutting-edge vision models: a Vmamba branch for fast and efficient long-range spatial feature extraction using State Space Models (SSMs), and a Vision Transformer (ViT) branch leveraging a pretrained backbone for deep semantic awareness through global self-attention. The complementary feature sets from these pathways are intelligently integrated using a multi-head cross-attention fusion mechanism, and the fused representation is passed to a regression head to predict a continuous beauty score. Evaluated on the SCUT-FBP5500 dataset using a rigorous 5-fold cross-validation protocol, our model achieves state-of-the-art performance with a Pearson Correlation (PC) of 0.9261. These results demonstrate the superior capability of our hybrid approach in capturing the complex perceptual cues essential for robust facial beauty assessment. The code is available at https://github.com/DjameleddineBoukhari/Hybrid-Vmamba-ViT-FBP.