计算机与现代化

• 图像处理 • 上一篇    下一篇

基于余弦相似度的边界样本选择方法

  

  1. 中国民航大学计算机科学与技术学院,天津300300
  • 收稿日期:2017-01-03 出版日期:2017-08-31 发布日期:2017-09-01
  • 作者简介:李春利(1964-),男,黑龙江哈尔滨人,中国民航大学计算机科学与技术学院副教授,博士,研究方向:模式识别; 柳振东(1992-),男,硕士研究生,研究方向:模式识别; 惠康华(1982-),男,讲师,博士,研究方向:模式识别。
  • 基金资助:
    中国民航大学科研启动基金资助项目(2010QD10X)

Boundary Sample Selection Method Based on Cosine Similarity

  1. School of Computer Science and Technology, Civil Aviation University of China, Tianjin 300300, China
  • Received:2017-01-03 Online:2017-08-31 Published:2017-09-01

摘要: 卷积神经网络模型的训练通常需要大量的训练样本,导致训练时间过长。针对这一问题,本文提出一种基于余弦相似度的边界样本选择方法,选取边界样本构造训练集。通过该方法分别对MNIST,CIFAR10,SVHN数据集进行样本选择,利用卷积神经网络分类器进行实验研究。实验结果表明:该方法能够保留训练集中的典型样本,剔除冗余样本,从而减少训练样本的数量,缩短网络训练时间,提高网络学习效率。

关键词:  , 深度学习, 卷积神经网络, 模式识别, 边界数据, 图像识别, 样本选择

Abstract:
Abstract: The training of convolution neural network usually requires a lot of training samples, which causes the training time be too long. To solve this problem, this paper presents a boundary sample selection method based on cosine similarity. We select boundary samples as the training set of convolution neural network, and carry out example selection experiment on the MNIST, CIFAR10 and SVHN data sets. Then a convolutional neural network is used to carry out experiments. Experimental results show that this method can preserve the typical samples in the training set and eliminate redundant samples. Thereby, the number of training samples is reduced, the network training time is shortened and the learning efficiency of network is improved.

Key words: deep learning, convolutional neural network, pattern recognition, boundary data, image recognition, sample selection

中图分类号: