Scholarlink

<p>Long-tailed image classification is a critical challenge in deep learning. Recently, Probabilistic Contrastive Learning has shown great potential for addressing this issue. By modeling feature vectors using the von Mises-Fisher distribution, ProCo effectively exploits the sparse and unevenly distributed samples of tail classes. However, when the backbone network fails to capture fine-grained features of tail classes, the inter-class discriminability of the model becomes limited. To tackle this limitation, we propose an Adaptive Feature Aggregation (AFA) strategy to enhance the feature modeling capacity of the ProCo backbone in long-tailed scenarios. The proposed AFA strategy consists of three key components: (1) Fine-grained feature reconstruction emphasizes semantically relevant features and suppresses redundant activations through spatial and channel optimization, thereby improving intra-class compactness and inter-class separability; (2) Multi-scale semantic capture integrates local and global contextual information via multi-scale convolutional structures, alleviating the limitations of single-scale modeling; (3) Adaptive multi-scale fusion dynamically selects optimal feature combinations based on sample characteristics, thus strengthening the robustness and generalization of tail class representations. By reinforcing feature discriminability and employing vMF-based class-conditional probability modeling, the proposed AFA-ProCo framework effectively compensates for the feature sparsity of tail classes within analytical contrastive optimization. Experimental results on standard long-tailed benchmarks datasets including CIFAR-10-LT, CIFAR-100-LT, and TinyImageNet-LT demonstrate that AFA-ProCo significantly improves the accuracy of tail classes while maintaining competitive head-class performance, outperforming existing re-sampling, re-weighting, and contrastive learning approaches.</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

Long-tailed image classification via adaptive feature aggregation in probabilistic contrastive framework

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Qiangkui Leng,
  • Chengjun Diao

摘要

Long-tailed image classification is a critical challenge in deep learning. Recently, Probabilistic Contrastive Learning has shown great potential for addressing this issue. By modeling feature vectors using the von Mises-Fisher distribution, ProCo effectively exploits the sparse and unevenly distributed samples of tail classes. However, when the backbone network fails to capture fine-grained features of tail classes, the inter-class discriminability of the model becomes limited. To tackle this limitation, we propose an Adaptive Feature Aggregation (AFA) strategy to enhance the feature modeling capacity of the ProCo backbone in long-tailed scenarios. The proposed AFA strategy consists of three key components: (1) Fine-grained feature reconstruction emphasizes semantically relevant features and suppresses redundant activations through spatial and channel optimization, thereby improving intra-class compactness and inter-class separability; (2) Multi-scale semantic capture integrates local and global contextual information via multi-scale convolutional structures, alleviating the limitations of single-scale modeling; (3) Adaptive multi-scale fusion dynamically selects optimal feature combinations based on sample characteristics, thus strengthening the robustness and generalization of tail class representations. By reinforcing feature discriminability and employing vMF-based class-conditional probability modeling, the proposed AFA-ProCo framework effectively compensates for the feature sparsity of tail classes within analytical contrastive optimization. Experimental results on standard long-tailed benchmarks datasets including CIFAR-10-LT, CIFAR-100-LT, and TinyImageNet-LT demonstrate that AFA-ProCo significantly improves the accuracy of tail classes while maintaining competitive head-class performance, outperforming existing re-sampling, re-weighting, and contrastive learning approaches.