Scholarlink

<p>It is essential for fine-grained visual classification (FGVC) tasks to explore and mine subtle and unique features between visually similar sub-categories. Most existing methods are dedicated to extracting more discriminative features or as many different features as possible to improve classification performance. However, there has been insufficient research into addressing misclassification (category confusion) issues between visually similar categories. In this paper, we propose an improved cross-entropy (ICE) loss function to avoid classification errors by recalculating the probability distribution of the top-k categories. In addition, we propose two modules to improve performance, a generic object localization module (OLM) that incorporates the connectivity component algorithm to acquire object instances, and a Cross-Fusion Attention Module (CFAM) to model spatial long-range dependencies. Our proposed method can be trained end-to-end and does not require bounding boxes/part annotations. Extensive experiments achieve competitive performance on the CUB-200-2011, FGVC Aircraft, and Stanford Cars datasets. Our code is public at <a href="https://github.com/satuoqaq/ACC-FGVC">https://github.com/satuoqaq/ACC-FGVC</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

Alleviating category confusion in fine-grained visual classification

  • 期刊论文

  • 发表日期：

    2025年2月27日

查看全文
  • Die Yu,
  • Zhaoyan Fang,
  • Yong Jiang

摘要

It is essential for fine-grained visual classification (FGVC) tasks to explore and mine subtle and unique features between visually similar sub-categories. Most existing methods are dedicated to extracting more discriminative features or as many different features as possible to improve classification performance. However, there has been insufficient research into addressing misclassification (category confusion) issues between visually similar categories. In this paper, we propose an improved cross-entropy (ICE) loss function to avoid classification errors by recalculating the probability distribution of the top-k categories. In addition, we propose two modules to improve performance, a generic object localization module (OLM) that incorporates the connectivity component algorithm to acquire object instances, and a Cross-Fusion Attention Module (CFAM) to model spatial long-range dependencies. Our proposed method can be trained end-to-end and does not require bounding boxes/part annotations. Extensive experiments achieve competitive performance on the CUB-200-2011, FGVC Aircraft, and Stanford Cars datasets. Our code is public at https://github.com/satuoqaq/ACC-FGVC.