Few-shot learning aims to transfer training models for base classes with sufficient labeled data to adapt to novel classes with only a few training examples. In this paper, we propose a deep search based prototype estimation method to improve few-shot learning. Based on the deep search, the similarity between the data in the base class can be used to better complete the prototype of the support set category. After obtaining similar samples, we can generate a prototype and then use it to estimate the distribution of the category to improve the few-shot learning task. A comparative experiment is carried out on three benchmark datasets to evaluate the effectiveness of our algorithm for few-shot learning. The results show that our method can generate more robust samples and significantly improve the few-shot learning, which is superior to the most advanced methods on these datasets.