计算机工程 ›› 2019, Vol. 45 ›› Issue (4): 211-216.doi: 10.19678/j.issn.1000-3428.0050129

• 人工智能及识别技术 • 上一篇    下一篇

卷积神经网络池化方法研究

周林勇1,2,谢晓尧1,刘志杰1,任笔墨2   

  1. 1.贵州师范大学 贵州省信息与计算科学重点实验室,贵阳 550001; 2.贵州财经大学 数学与统计学院,贵阳 550025
  • 收稿日期:2018-01-16 出版日期:2019-04-15 发布日期:2019-04-15
  • 作者简介:周林勇(1987—),男,博士研究生,主研方向为深度学习、图像处理;谢晓尧(通信作者),教授、博士生导师;刘志杰,教授、博士;任笔墨,硕士。
  • 基金项目:

    国家自然科学基金(U1631132)。

Research on Pooling Method of Convolution Neural Network

ZHOU Linyong1,2,XIE Xiaoyao1,LIU Zhijie1,REN Bimo2   

  1. 1.Key Laboratory of Information and Computing Science of Guizhou Province,Guizhou Normal University,Guiyang 550001,China; 2.School of Mathematics and Statistics,Guizhou University of Finance and Economics,Guiyang 550025,China
  • Received:2018-01-16 Online:2019-04-15 Published:2019-04-15

摘要:

为解决随机池化中零元素概率为0导致不能被选择的问题,提出一种改进的混合概率随机池化方法。将池化域中的元素去重复并按升序排序,然后加上对应次序的幂次,得到元素的权重概率。在此基础上,根据多项分布取样给出池化值。在数据集MNIST、CIFAR-10、CIFAR-100上进行实验,结果表明,该方法在3种数据集上的分类准确率分别为99.50%、72.25%、39.05%,相较于传统池化方法具有较好的分类效果与稳健性。

关键词: 卷积神经网络, 深度学习, 池化方法, 多项分布, 图像分类

Abstract:

In order to solve the problem that the probability of zero elements in random pooling is zero,it cannot be selected.An improved hybrid probability stochastic method is proposed.The elements in the pooled domain are deduplicated and sorted in ascending order,and then the power of the corresponding order is added to obtain the weight probability of the element.On this basis,the pooling value is given based on the multi-distribution sampling.Experimental results on the datasets MNIST,CIFAR-10,and CIFAR-100 show that the method in the classification accuracy of the three datasets is 99.50%,72.25%,and 39.05%,respectively.Compared with the traditional pooling method,the method has good classification effect and robustness.

Key words: Convolutional Neural Network(CNN), deep learning, pooling method, multinomial distribution, image classification

中图分类号: