基于新型初始模块的卷积神经网络图像分类方法

doi:10.16180/j.cnki.issn1007-7820.2021.02.009

摘要/Abstract

摘要：

在涉及分类识别的问题中,首选方法是基于卷积神经网络的分类方法。为解决传统卷积神经网络处理能力较差、分类精度较低等问题,文中提出了一种新型的卷积神经网络图像分类模型。一方面在传统的网络模型基础上增加新型Inception模块,增强了模型的特征信息的融合,提高了特征表达的能力;另一方面通过激活函数、数据增强、批量正则化、权重初始优化以及Adadelta优化方法来改善模型的性能,提升分类准确率。通过基于新型初始模块的新型网络模型对CIFAR-10数据集上的数据进行试验,并与传统网络模型方法对比,证明改进模型能有效提升网络性能。

关键词: 卷积神经网络, 图像分类, 新型Inception模块, 数据增强, 权重初始化, Adadelta优化方法

Abstract:

Among the problems involving image classification, the classification method based on convolutional neural networks is preferred. In order to solve the problems of poor processing capacity and low classification accuracy, a new type of image classification model of convolutional neural network is proposed. The new Inception module is added on the basis of the traditional network model, which enhances the transmission of the feature information of the model and improves the ability of feature expression. Additionally, the performance of the model and the classification accuracy are improved by activation function, data enhancement, batch regularization, initial weight optimization and Adadelta optimization method. The new network model based on the new initial module is used to test the data on the CIFAR-10 data set, and compared with the traditional network model method, which proves that the modulated model effectives improves the network performance.

Key words: convolution neural network, image classification, new Inception module, data augmentation, weight initialization, Adadelta optimization method

中图分类号:

TN391.41

朱斌,刘子龙. 基于新型初始模块的卷积神经网络图像分类方法[J]. 电子科技, 2021, 34(2): 52-56.

ZHU Bin,LIU Zilong. Image Classification Method Using Convolutional Neural Network Based on New Initial Module[J]. Electronic Science and Technology, 2021, 34(2): 52-56.

图/表 12

图1

图2

图3

图4

表1

图5

图6

图7

图8

图9

图10

表2

参考文献 16

[1]	Zeng Y, Xu X, Fang Y, et al. Traffic sign recognition using deep convolutional networks and extreme learning machine [C].Helsinki:International Conference on Intelligent Science and Big Data, 2015.
[2]	Weng Q, Mao Z, Lin J, et al. Land-use classification via extreme learning classifier based on deep convolutional features[J]. IEEE Geoscience and Remote Sensing Letters, 2017,14(5):704-708.
[3]	Huang G B, Zhu Q Y, Siew C K. Extreme learning machine:theory and applications[J]. Neurocomputing, 2006,70(1-3):489-501. doi: 10.1016/j.neucom.2005.12.126
[4]	Redmon J, Divvala S, Girshick R, et al. You only look once:unified, real-time object detection [C].Las Vegas:IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[5]	熊海朋, 陈洋洋, 陈春玮. 基于卷积神经网络的场景图像文本定位研究[J]. 电子科技, 2018,31(1):50-53.
	Xiong Haipeng, Chen Yangyang, Chen Chunwei. Text localization in image based on convolutional neural network[J]. Electronic Science and Technology, 2018,31(1):50-53.
[6]	王丽君, 于莲芝. 基于卷积神经网络的位置识别[J]. 电子科技, 2017,30(1):104-106.
	Wang Lijun, Yu Lianzhi. Visual place recognition based on convolutional neural network[J]. Electronic Science and Technology, 2017,30(1):104-106.
[7]	Hinton G E, Salakhutdinov R R. Reducing the dimensionality of data with neural networks[J]. Science, 2016,313(5786):504-507. doi: 10.1126/science.1127647 pmid: 16873662
[8]	Kingma D, Ba J. ADAM:a method for stochastic optimization[J]. Computer Science, 2014,26(2):358-366.
[9]	白琮, 黄玲, 陈佳楠. 面向大规模图像分类深度卷积神经网络优化[J]. 软件学报, 2018,29(4):1029-1038.
	Bai Cong, Huang Ling, Chen Jianan. Optimization of deep convolutional neural networks for large-scale image classification[J]. Journal of Software, 2018,29(4):1029-1038.
[10]	奚雪峰, 周国栋. 面向自然语言处理的深度学习研究[J]. 自动化学报, 2016,42(10):1445-1465. doi: 10.16383/j.aas.2016.c150682
	Xi Xuefeng, Zhou Guodong. A survey on deep learning for natural language processing[J]. Acta Automatica Sinica, 2016,42(10):1445-1465. doi: 10.16383/j.aas.2016.c150682
[11]	Rong H J, Ong Y S, Tan A H, et al. A fast pruned-extreme learning machine for classification problem[J]. Neurocomputing, 2008,72(1-3):359-366.
[12]	赵亮, 王晓峰. 基于深度卷积神经网络的船舶识别方法[J]. 舰船科学技术, 2016,38(15):119-123.
	Zhao Liang, Wang Xiaofeng. Research on ship recognition method based on deep convolutional neural network[J]. Ship Science and Technology, 2016,38(15):119-123.
[13]	Hinton G. A practical guide to training restricted boltzmann machines[J]. Momentum, 2010,9(1):926-947.
[14]	张振焕, 周彩兰. 基于残差优化卷积神经网络服装分类[J]. 计算机工程与科学, 2018,40(2):354-360.
	Zhang Zhenhuan, Zhou Cailan. An optimized clothing classification algorithm based on residual convolutional neural network[J]. Computer Engineering and Science, 2018,40(2):354-360.
[15]	李鸣, 张鸿. 基于卷积神经网络迭代优化图像分类算法[J]. 计算机工程与设计, 2017,38(1):198-202.
	Li Ming, Zhang Hong. Image classification based on convolutional neural network of iterative optimization[J]. Computer Engineering and Design, 2017,38(1):198-202.
[16]	易超人, 邓燕妮. 多通道卷积神经网络图像识别方法[J]. 河南科技大学学报, 2017,38(3):41-44.
	Yi Chaoren, Deng Yanni. Image recognition method of multi-channel convolutional neural networks[J]. Journal of Henan University of Science and Technology(Natural Science), 2017,38(3):41-44.

网络层	类别	卷积核	步长	深度	激活函数
Input	输入层	—	—	3	ReLU
Conv1	卷积层	3×3	1	64	ReLU
Maxpool2	池化层	2×2	2	64	ReLU
Conv3	卷积层	3×3	1	128	ReLU
Maxpool4	池化层	2×2	2	128	ReLU
Conv5	卷积层	3×3	1	256	ReLU
Maxpool6	池化层	2×2	2	256	ReLU
Inception7	初始层	1×1,3×3, 3×1,1×3	—	768	ReLU
FC8	全连接	—	—	—	ReLU+Dropout
FC9	全连接	—	—	—	ReLU+Dropout
Output	输出	—	—	—	Softmax

模型结构	准确率
传统CNN	68.94%
传统CNN+ReLU+Adadelta+ Data Augmentation	74.28%
改进CNN	73.11%
改进CNN+ ReLU+Adadelta +Data Augmentation	84.23%