摘要：

为提高卷积神经网络的分类精度,提出一种结合多个网络的改进Stacking算法。将卷积神经网络作为基分类器对数据进行分类,得到新的样本再经过元分类器分类。为降低元分类器输入数据的维度和多个网络分类结果之间的相关性,采用主成分分析方法对基分类器的输出进行降维。在数据集上进行分类精度对比实验,结果表明,与传统Stacking、基于平均后验概率算法和基于类投票算法相比,该算法在同类型网络和不同类型网络中,分类精度均较高且更具有稳定性。

关键词: 卷积神经网络, Stacking算法, 主成分分析, 降维, 网络结构, 分类精度

Abstract:

In order to improve the classification accuracy of convolution neural network,an improved Stacking algorithm combining multiple convolution neural networks is proposed.The convolution neural network is used as the base classifier to classify the data,and the new sample is then classified by the meta-classifier.In order to reduce the dimension and correlation of the input data in meta-classifier,the dimension of output data of base classifier are reducted by Principal Component Analysis(PCA).Experimental results on classification accuracy of data sets show that,compared with traditional Stacking,average posteriori probability based algorithm and class voting based algorithm,the classification accuracy of the proposed algorithm is higher and more stable in similar networks and different types of networks.

Key words: convolution neural network, Stacking algorithm, Principal Component Analysis(PCA), dimensionality reduction, network structure, classification accuracy

中图分类号: 

• TP391