Scholarlink

<p>Fine-grained visual classification (FGVC) poses significant challenges due to subtle inter-class differences and large intra-class variations. Addressing these issues requires capturing global features of target regions while accurately extracting detailed local features. In this paper, we propose a fine-grained visual classification network based on curriculum learning and global–local feature interaction (CLFI-Net) to effectively enhance the performance of FGVC. The network employs a feature enhancement module to improve the input quality of part navigator, enabling precise localization of key regions while mitigating the impact of complex backgrounds. A global–local feature interaction module is designed to capture the relationship between global and local features, suppressing redundant information within bounding boxes. Furthermore, an enhanced multi-stage curriculum learning module optimizes feature representations at each stage using a multi-scale channel attention mechanism. Combined with a label smoothing strategy, it progressively guides the learning of discriminative features from shallow to deep layers. We conduct extensive evaluations of CLFI-Net on three popular FGVC datasets and a self-constructed algae dataset. Extensive evaluations on three benchmark FGVC datasets and a self-constructed algae dataset demonstrate that CLFI-Net achieves superior classification performance, validating its effectiveness and robustness. Code is available at: <a href="https://github.com/xiazi777/CLFI-Net.">https://github.com/xiazi777/CLFI-Net.</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 Fine-Grained Visual Classification via Curriculum Learning and Global–Local Feature Interaction

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Xueqing Zhang,
  • Shuo Wang,
  • Fengjuan Feng,
  • Jianlei Liu

摘要

Fine-grained visual classification (FGVC) poses significant challenges due to subtle inter-class differences and large intra-class variations. Addressing these issues requires capturing global features of target regions while accurately extracting detailed local features. In this paper, we propose a fine-grained visual classification network based on curriculum learning and global–local feature interaction (CLFI-Net) to effectively enhance the performance of FGVC. The network employs a feature enhancement module to improve the input quality of part navigator, enabling precise localization of key regions while mitigating the impact of complex backgrounds. A global–local feature interaction module is designed to capture the relationship between global and local features, suppressing redundant information within bounding boxes. Furthermore, an enhanced multi-stage curriculum learning module optimizes feature representations at each stage using a multi-scale channel attention mechanism. Combined with a label smoothing strategy, it progressively guides the learning of discriminative features from shallow to deep layers. We conduct extensive evaluations of CLFI-Net on three popular FGVC datasets and a self-constructed algae dataset. Extensive evaluations on three benchmark FGVC datasets and a self-constructed algae dataset demonstrate that CLFI-Net achieves superior classification performance, validating its effectiveness and robustness. Code is available at: https://github.com/xiazi777/CLFI-Net.