Scholarlink

<p>Kinship verification plays a pivotal role in security monitoring, genetic research, and social media analysis. Traditional CNN-based approaches struggle with limited receptive fields, impeding their ability to model long-range dependencies. Vision transformer, on the other hand, excels at capturing global context but often overlook fine-grained details. To address these limitations, we propose the context-aware multi-scale transformer (CAMS-Trans), a hybrid framework that integrates the inductive bias of convolutions with the global context modeling capabilities of transformers. Our approach comprises two core modules: the multi-scale context embedding (MSCE) module, which utilizes dilated convolutions to generate hierarchical features capturing fine-grained textures and coarse-grained contours; and the global–local integration encoder (GLIE) module, which dynamically injects local features extracted by a lightweight convolutional branch into transformer layers. Experiments on the FIW and KinFaceW datasets demonstrate that CAMS-Trans achieves state-of-the-art performance, outperforming existing methods with average accuracies of 81.32% on FIW, 84.70% on KinFaceW-I, and 93.25% on KinFaceW-II. Here, we show that the collaborative design of MSCE and GLIE enables effective kinship feature learning, balancing local details and global context. The code is publicly available on <a href="https://github.com/BoyuanL/MSCA-msvit/tree/master">https://github.com/BoyuanL/MSCA-msvit/tree/master</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

Enhanced kinship verification via context-aware multi-scale transformer

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Xiaoke Zhu,
  • Boyuan Li,
  • Xiaopan Chen,
  • Fumin Qi,
  • Caihong Yuan,
  • Xiao-Yuan Jing

摘要

Kinship verification plays a pivotal role in security monitoring, genetic research, and social media analysis. Traditional CNN-based approaches struggle with limited receptive fields, impeding their ability to model long-range dependencies. Vision transformer, on the other hand, excels at capturing global context but often overlook fine-grained details. To address these limitations, we propose the context-aware multi-scale transformer (CAMS-Trans), a hybrid framework that integrates the inductive bias of convolutions with the global context modeling capabilities of transformers. Our approach comprises two core modules: the multi-scale context embedding (MSCE) module, which utilizes dilated convolutions to generate hierarchical features capturing fine-grained textures and coarse-grained contours; and the global–local integration encoder (GLIE) module, which dynamically injects local features extracted by a lightweight convolutional branch into transformer layers. Experiments on the FIW and KinFaceW datasets demonstrate that CAMS-Trans achieves state-of-the-art performance, outperforming existing methods with average accuracies of 81.32% on FIW, 84.70% on KinFaceW-I, and 93.25% on KinFaceW-II. Here, we show that the collaborative design of MSCE and GLIE enables effective kinship feature learning, balancing local details and global context. The code is publicly available on https://github.com/BoyuanL/MSCA-msvit/tree/master.