Scholarlink

<p>In practical applications, datasets often exhibit imbalanced distributions. Deep models trained on imbalanced data frequently produce biased results, leading to degraded generalization performance. Data augmentation techniques represent a key approach to addressing imbalanced data. Although random image class mixing is widely regarded as a reliable method, existing mix-based augmentation strategies often neglect the spatial relationships between majority and minority class images, potentially generating low-quality samples. In this paper, we adopt an optimal transport approach to perform pre-matching based on the spatial relationships between primary objects across images before class mixing. This method ensures the accuracy of primary object positioning for both majority and minority classes while maximizing the utilization of majority-class background images. Experimental results demonstrate that as a data-centric input-level optimization method, our approach achieves superior performance across various imbalanced scenarios and exhibits scalability.</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

Positional Relation Contextual Mixing for Imbalanced Classification

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Yucheng Jiang,
  • Jiateng Li,
  • Yuan Tian,
  • Jiangchao Yao,
  • Xin Yu,
  • Wei Ye,
  • Xiaofeng Cao

摘要

In practical applications, datasets often exhibit imbalanced distributions. Deep models trained on imbalanced data frequently produce biased results, leading to degraded generalization performance. Data augmentation techniques represent a key approach to addressing imbalanced data. Although random image class mixing is widely regarded as a reliable method, existing mix-based augmentation strategies often neglect the spatial relationships between majority and minority class images, potentially generating low-quality samples. In this paper, we adopt an optimal transport approach to perform pre-matching based on the spatial relationships between primary objects across images before class mixing. This method ensures the accuracy of primary object positioning for both majority and minority classes while maximizing the utilization of majority-class background images. Experimental results demonstrate that as a data-centric input-level optimization method, our approach achieves superior performance across various imbalanced scenarios and exhibits scalability.