计算机与现代化

• 图像处理 • 上一篇    下一篇

基于原型网络的小样本图像识别方法

  

  1. (上海理工大学光电信息与计算机工程学院,上海200093)
  • 收稿日期:2019-07-08 出版日期:2020-03-24 发布日期:2020-03-30
  • 作者简介:樊笛(1994-),男,河南郑州人,硕士研究生,研究方向:模式识别与机器视觉,E-mail: 1219045134@qq.com; 巨志勇(1975-),男,山东济南人,讲师,博士,研究方向:图像处理与模式识别,E-mail: juzy@usst.edu.cn。
  • 基金资助:
    国家自然科学基金资助项目(81101116)

Method of Small Sample Image Recognition Based on Prototype Network

  1. (School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China)
  • Received:2019-07-08 Online:2020-03-24 Published:2020-03-30

摘要: 当前的图像识别领域,大部分的分类或者识别方法都建立在已有大量数据的基础上,将大量数据投入训练,经过采样分析、特征提取后做判别分类。然而在现实世界中,大多数目标分类问题并没有大量的标注数据。为了解决基于小样本数据集的图像识别问题,本文首先使用数据增强方法扩充数据集,然后利用多层卷积神经网络将图像映射到高维嵌入空间中,再使用原型网络得到每个类的原型点,根据嵌入空间中测试图像与各个类原型点之间的距离将其分类。实验结果表明,该方法在小样本条件下具有较高的识别准确率和较强的鲁棒性。

关键词: 图像识别, 小样本, 卷积神经网络, 原型网络, 嵌入空间, 数据增强

Abstract: In the current image recognition field, most of the classification or recognition methods are built on the basis of existing large amounts of data, which are put into training and classified through sampling analysis and feature extraction. However, in the real world, most target classification problems do not have a large amount of annotated data. In order to solve the problem of image recognition based on small data sets, this paper uses the data augmentation to enhance data sets, and uses multi-layer CNN to map the image into high-dimensional space, then gets prototypes of each class by using the prototype network. Finally, the test image can be classified according to the distance among prototype points and test image in the embedded space. Experimental results show that this method has high recognition accuracy under the condition of small data set, and also has good stability and strong robustness.

Key words: image recognition, small data set, CNN, prototype network, embedding space, data augmentation

中图分类号: