Scholarlink

<p>In recent years, researchers have developed various classification methods using feature reconstruction to reduce subtle differences between categories in few-shot fine-grained image classification (FS-FGIC). However, existing methods over-rely on deep semantic features while neglecting fine-grained shallow details, resulting in an imbalance between deep and shallow representations. Additionally, traditional multilayer feature fusion methods are prone to background noise and struggle to focus dynamically on discriminative regions. To address this, we propose the multilayer feature dynamic interaction fusion network (MDF-Net). Specifically, the multilayer feature interaction (MFI) module adaptively fuses shallow local details with deep semantic features using cross-scale covariance correlation computation and residual-guided bidirectional attention propagation. Meanwhile, the dynamic feature fusion (DFF) module employs an adaptive weighting mechanism to aggregate multilayer information, suppress background noise, and enhance gradient-sensitive features in discriminative regions. Extensive experiments validate the competitive performance of our method on benchmark datasets. The code will be available at:&#xa0;<a href="https://github.com/CYu-stu/MDF.git">https://github.com/CYu-stu/MDF.git</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

MDF-Net: multilayer feature dynamic interactive fusion network for few-shot fine-grained image classification

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Yu Cheng,
  • Bo Li,
  • Penghao Jia

摘要

In recent years, researchers have developed various classification methods using feature reconstruction to reduce subtle differences between categories in few-shot fine-grained image classification (FS-FGIC). However, existing methods over-rely on deep semantic features while neglecting fine-grained shallow details, resulting in an imbalance between deep and shallow representations. Additionally, traditional multilayer feature fusion methods are prone to background noise and struggle to focus dynamically on discriminative regions. To address this, we propose the multilayer feature dynamic interaction fusion network (MDF-Net). Specifically, the multilayer feature interaction (MFI) module adaptively fuses shallow local details with deep semantic features using cross-scale covariance correlation computation and residual-guided bidirectional attention propagation. Meanwhile, the dynamic feature fusion (DFF) module employs an adaptive weighting mechanism to aggregate multilayer information, suppress background noise, and enhance gradient-sensitive features in discriminative regions. Extensive experiments validate the competitive performance of our method on benchmark datasets. The code will be available at: https://github.com/CYu-stu/MDF.git.