Scholarlink

<p>To address poor accuracy and inter-class confusion caused by tail-class sparsity and overlapping distributions in long-tailed data, this paper proposes a Distribution Correlation Weighting and Confusion-Aware Multi-Modal Enhancement (DCW-CAME) algorithm. A Distribution Correlation Weighting (DCW) mechanism based on simplified Wasserstein distance constructs a Distribution Correlation Index (DCI) to quantify distribution overlap and adapt loss weights. Then, a Confusion-Aware Multi-Modal Enhancement (CAME) module jointly performs morphological, contextual, and gradient-based augmentations to enrich feature diversity and generalization for confusing classes. Finally, a confusion-aware boundary calibration imposes adaptive penalties via the Overlap Region Error (ORE) to refine decision boundaries. Experiments on CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 show that DCW-CAME outperforms classical methods such as Cross-Entropy and LDAM, achieving 4.92% improvement over Cross-Entropy on CIFAR-100-LT (<i>λ </i>= 200). Visualization results confirm clearer feature separation and reduced confusion, validating DCW-CAME’s effectiveness for long-tailed classification.</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

DCW-CAME: distribution correlation weighting and confusion-aware multi-modal enhancement for long-tailed classification

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Xu Lian,
  • Zhen Xue,
  • Liangliang Zhang,
  • Tiantian Li

摘要

To address poor accuracy and inter-class confusion caused by tail-class sparsity and overlapping distributions in long-tailed data, this paper proposes a Distribution Correlation Weighting and Confusion-Aware Multi-Modal Enhancement (DCW-CAME) algorithm. A Distribution Correlation Weighting (DCW) mechanism based on simplified Wasserstein distance constructs a Distribution Correlation Index (DCI) to quantify distribution overlap and adapt loss weights. Then, a Confusion-Aware Multi-Modal Enhancement (CAME) module jointly performs morphological, contextual, and gradient-based augmentations to enrich feature diversity and generalization for confusing classes. Finally, a confusion-aware boundary calibration imposes adaptive penalties via the Overlap Region Error (ORE) to refine decision boundaries. Experiments on CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 show that DCW-CAME outperforms classical methods such as Cross-Entropy and LDAM, achieving 4.92% improvement over Cross-Entropy on CIFAR-100-LT (λ = 200). Visualization results confirm clearer feature separation and reduced confusion, validating DCW-CAME’s effectiveness for long-tailed classification.