Scholarlink

<p>In the context of limited data samples, few-shot learning continues to pose a significant challenge. A prevalent strategy in recent times has been to pre-train models on extensive datasets and subsequently transfer them to downstream few-shot tasks, which has demonstrated efficacy in enhancing performance. However, a persistent challenge lies in the inadequacy of pre-trained models to capture the essential features of the new downstream dataset. This issue is particularly acute in images containing multiple entities, where crucial features are often overlooked yet play a pivotal role in image classification. To address this challenge, we propose an innovative local information enhancement strategy that harnesses information from all samples to capture important local features in images and integrates them with global features. The objective of our strategy is to enhance class differentiation by ensuring distinct class prototypes in the embedding space through the incorporation of local information. By integrating local information, query samples exhibit closer alignment with the prototype of their respective classes, ultimately resulting in improved classification accuracy. To further bolster the performance of few-shot classification, we have refined the pre-trained model approach and augmented the dataset. Comprehensive ablation experiments demonstrate the specific impact of our approach on enhancing the accuracy of few-shot 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

Boosting few-shot learning via selective patch embedding by comprehensive sample analysis

  • 期刊论文

  • 发表日期：

    2025年2月20日

查看全文
  • Juan Yang,
  • Yuliang Zhang,
  • Ronggui Wang,
  • Lixia Xue

摘要

In the context of limited data samples, few-shot learning continues to pose a significant challenge. A prevalent strategy in recent times has been to pre-train models on extensive datasets and subsequently transfer them to downstream few-shot tasks, which has demonstrated efficacy in enhancing performance. However, a persistent challenge lies in the inadequacy of pre-trained models to capture the essential features of the new downstream dataset. This issue is particularly acute in images containing multiple entities, where crucial features are often overlooked yet play a pivotal role in image classification. To address this challenge, we propose an innovative local information enhancement strategy that harnesses information from all samples to capture important local features in images and integrates them with global features. The objective of our strategy is to enhance class differentiation by ensuring distinct class prototypes in the embedding space through the incorporation of local information. By integrating local information, query samples exhibit closer alignment with the prototype of their respective classes, ultimately resulting in improved classification accuracy. To further bolster the performance of few-shot classification, we have refined the pre-trained model approach and augmented the dataset. Comprehensive ablation experiments demonstrate the specific impact of our approach on enhancing the accuracy of few-shot classification.