Scholarlink

<p>Learning discriminative features is crucial for deep learning-based image classification, particularly in open-set scenarios. Current methods often aim to achieve this by increasing inter-class decision boundaries and compressing intra-class feature distances through the addition of a fixed margin to all samples. However, assigning the same margin to all samples disregards the varying learning difficulties among different samples. Moreover, label noise is almost unavoidable in large-scale classification datasets. To address these issues, we devise a novel loss function that adaptively adjusts the angular margin based on the difficulty of each sample. Specifically, this paper proposes a quantification function that helps the model identify sample difficulty during training and adaptively adjusts the angular margin for different samples based on these quantification results. Additionally, we propose an automatic detection method for anomalous labels during training, leveraging the Laplace kernel function, which significantly enhances model performance in the presence of noisy label data. To verify the effectiveness of the proposed method, we conduct several experiments using face recognition tasks and typical open-set image recognition as case studies, demonstrating that our method outperforms the state-of-the-art approaches. The source code is available at: <a href="https://github.com/TCCofWANG/Hardness-Aware-Angular-Margin-Loss">https://github.com/TCCofWANG/Hardness-Aware-Angular-Margin-Loss</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

An open-set image classification algorithm based on hardness-aware angular margin loss and label noise filtering

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Jinpeng Cao,
  • Hao Wang,
  • Wei Zhang,
  • Shichao Ren,
  • Chi-Sing Leung

摘要

Learning discriminative features is crucial for deep learning-based image classification, particularly in open-set scenarios. Current methods often aim to achieve this by increasing inter-class decision boundaries and compressing intra-class feature distances through the addition of a fixed margin to all samples. However, assigning the same margin to all samples disregards the varying learning difficulties among different samples. Moreover, label noise is almost unavoidable in large-scale classification datasets. To address these issues, we devise a novel loss function that adaptively adjusts the angular margin based on the difficulty of each sample. Specifically, this paper proposes a quantification function that helps the model identify sample difficulty during training and adaptively adjusts the angular margin for different samples based on these quantification results. Additionally, we propose an automatic detection method for anomalous labels during training, leveraging the Laplace kernel function, which significantly enhances model performance in the presence of noisy label data. To verify the effectiveness of the proposed method, we conduct several experiments using face recognition tasks and typical open-set image recognition as case studies, demonstrating that our method outperforms the state-of-the-art approaches. The source code is available at: https://github.com/TCCofWANG/Hardness-Aware-Angular-Margin-Loss.