Scholarlink

<p>The few-shot classification task is a challenging task in which there are only a few samples for each new category. Traditional classifiers often suffer from overfitting when facing such tasks, because a small amount of data does not reflect the true distribution of a large amount of data well, leading to a decrease in classification accuracy. To address the above challenges, this study designs a prototype complementation and dynamic loss adjustment based few-shot reconstruction network (PCDL-FSCN). We propose a prototype complementation strategy that augments the support set through sample augmentation and expands its effective sample size. This approach incorporates a dynamic loss function which reduces positive sample loss while enhancing negative sample loss, thereby enabling the model to focus on hard-to-classify samples. After extensive testing, the 1-shot and 5-shot learning performance of PCDL-FSCN on the mini-ImageNet dataset reached 72.92% and 84.91%, respectively; On tiered-ImageNet, the rates were 77.89% and 87.72%, respectively, demonstrating their effectiveness.</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

Prototype completion and dynamic loss in few-shot classification network

  • 期刊论文

  • 发表日期：

    2025年7月18日

查看全文
  • Jinhu Wu,
  • Junyang Yu,
  • Shiqian Wang,
  • Han Li,
  • Zhicheng Wang,
  • Tingyu Wang

摘要

The few-shot classification task is a challenging task in which there are only a few samples for each new category. Traditional classifiers often suffer from overfitting when facing such tasks, because a small amount of data does not reflect the true distribution of a large amount of data well, leading to a decrease in classification accuracy. To address the above challenges, this study designs a prototype complementation and dynamic loss adjustment based few-shot reconstruction network (PCDL-FSCN). We propose a prototype complementation strategy that augments the support set through sample augmentation and expands its effective sample size. This approach incorporates a dynamic loss function which reduces positive sample loss while enhancing negative sample loss, thereby enabling the model to focus on hard-to-classify samples. After extensive testing, the 1-shot and 5-shot learning performance of PCDL-FSCN on the mini-ImageNet dataset reached 72.92% and 84.91%, respectively; On tiered-ImageNet, the rates were 77.89% and 87.72%, respectively, demonstrating their effectiveness.