Scholarlink

<p>Few-shot learning (FSL) aims to achieve accurate recognition of unseen classes with extremely limited samples for each class. In recent years, numerous methods adopt prototype-based strategies to address the FSL challenges, leading to significant advancements in the field. However, this class of methods may suffer from prototype bias issues due to the presence of outlier samples in support set and the scarcity of samples. To address these issues, we propose a GAN-based prototype refinement network (GPRN) for few-shot learning. Specifically, to solve the outlier samples problem, we use a pre-correction module to selectively enhance the discriminability of sample features. Moreover, we employ a novel Conditional Wasserstein Generative Adversarial Network in the FSL framework to generate additional sample features for each class to solve the sample scarcity problem. And, by considering the relationship between generated sample features and class prototypes, different weight coefficients are assigned to each generated feature to reduce the interference from noisy features within the generated features. Finally, a prototype refinement module is used to fuse the few-shot prototypes and fake prototypes to obtain the final rectified prototypes. We show that the rectified prototypes are closer to the true class centers, and when these prototypes are used in subsequent classification measurements, the classification results are significantly improved. Experimental results on four benchmark datasets, <i>mini</i>-ImageNet, <i>tiered</i>-ImageNet, CIFAR-FS, and CUB-200–2011, show that our method achieves 5-way 1-shot accuracy of 53.42%, 55.74%, 62.40%, and 59.21%, and 5-way 5-shot accuracy of 69.42%, 72.75%, 78.58%, and 73.42%, respectively, demonstrating the validity of GPRN.</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

GPRN: GAN-based prototype refinement network for few-shot learning

  • 期刊论文

  • 发表日期：

    2025年5月7日

查看全文
  • Liufei Luo,
  • Zhao Yang,
  • Zhiyong Chen,
  • Renlong Cao,
  • Yusen Liu

摘要

Few-shot learning (FSL) aims to achieve accurate recognition of unseen classes with extremely limited samples for each class. In recent years, numerous methods adopt prototype-based strategies to address the FSL challenges, leading to significant advancements in the field. However, this class of methods may suffer from prototype bias issues due to the presence of outlier samples in support set and the scarcity of samples. To address these issues, we propose a GAN-based prototype refinement network (GPRN) for few-shot learning. Specifically, to solve the outlier samples problem, we use a pre-correction module to selectively enhance the discriminability of sample features. Moreover, we employ a novel Conditional Wasserstein Generative Adversarial Network in the FSL framework to generate additional sample features for each class to solve the sample scarcity problem. And, by considering the relationship between generated sample features and class prototypes, different weight coefficients are assigned to each generated feature to reduce the interference from noisy features within the generated features. Finally, a prototype refinement module is used to fuse the few-shot prototypes and fake prototypes to obtain the final rectified prototypes. We show that the rectified prototypes are closer to the true class centers, and when these prototypes are used in subsequent classification measurements, the classification results are significantly improved. Experimental results on four benchmark datasets, mini-ImageNet, tiered-ImageNet, CIFAR-FS, and CUB-200–2011, show that our method achieves 5-way 1-shot accuracy of 53.42%, 55.74%, 62.40%, and 59.21%, and 5-way 5-shot accuracy of 69.42%, 72.75%, 78.58%, and 73.42%, respectively, demonstrating the validity of GPRN.