Scholarlink

<p>Few-shot fine-grained image classification (FSFG) aims to recognize new fine-grained categories with few labeled samples. The rich local information contained in mid-level features is beneficial for fine-grained image classification, while most existing methods primarily rely on high-level features with global semantics and overlook mid-level features. Moreover, target objects with varying sizes lead to discriminative details distributed across different spatial scales. To address these challenges, we propose BFGL, a dual-branch framework for Boosting Few-Shot Fine-Grained Image Classification with Global and Local Feature Learning. We design a MultiLayer Attention Fusion (MLAF) module to adaptively integrate multi-level features and suppress background noise, enabling the network to fully exploit informative mid-level features. To capture local detailed features at multiple scales, we propose a Multi-Scale Feature Enhancement (MSFE) module. In addition, we develop a Cross-Sample Channel Alignment (CSCA) module that aligns channel-wise responses between support and query samples, thereby enhancing discriminative channels. Experimental results demonstrate superior performance across five standard fine-grained datasets.</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

Boosting few-shot fine-grained image classification with global and local feature learning

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Xuhao Shi,
  • Baolin Liu

摘要

Few-shot fine-grained image classification (FSFG) aims to recognize new fine-grained categories with few labeled samples. The rich local information contained in mid-level features is beneficial for fine-grained image classification, while most existing methods primarily rely on high-level features with global semantics and overlook mid-level features. Moreover, target objects with varying sizes lead to discriminative details distributed across different spatial scales. To address these challenges, we propose BFGL, a dual-branch framework for Boosting Few-Shot Fine-Grained Image Classification with Global and Local Feature Learning. We design a MultiLayer Attention Fusion (MLAF) module to adaptively integrate multi-level features and suppress background noise, enabling the network to fully exploit informative mid-level features. To capture local detailed features at multiple scales, we propose a Multi-Scale Feature Enhancement (MSFE) module. In addition, we develop a Cross-Sample Channel Alignment (CSCA) module that aligns channel-wise responses between support and query samples, thereby enhancing discriminative channels. Experimental results demonstrate superior performance across five standard fine-grained datasets.