基于数据增强的卷积神经网络图像识别研究  

Research on Image Recognition of Convolution Neural Network Based on Data Enhancement

作　　者：高友文 周本君 胡晓飞[2] GAO You-wen;ZHOU Ben-jun;HU Xiao-fei(School of Telecommunication ＆ Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Geography ＆ Bioinformatics,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学地理与生物信息学院,江苏南京210003

出　　处：《计算机技术与发展》2018年第8期62-65,共4页Computer Technology and Development

基　　金：国家自然科学基金(61271082);江苏省重点研发计划(BE2015700);江苏省自然科学基金(BK20141432)

摘　　要：针对深度学习网络在处理图像分类的过程中数据集样本数较少和样本相似度较高的问题,在卷积神经网络模型Alex Net的基础上,提出了对数据集采用数据集扩增、背景分割和主成分分析等数据预处理方法。卷积神经网络模型的基本结构为5个卷积层,2个全连接层和dropout层。实验环境是ubuntu16.04系统,Caffe深度学习框架。实验首先对原始的公开数据集Leaves和苹果表面病变数据集进行分类识别测试,分别得到84%和78%的准确率。然后对数据增强后的数据集再进行测试,公开数据集leaves的准确率为86%,准确率提高了2%,苹果表面病变数据集的准确率为83%,准确率提高了5%。测试结果表明,通过数据增强处理后,公开数据集Leaves和苹果表面病变数据集在该网络上的识别准确率都有了一定的提升。Aiming at the problem that the number of data sets is small and the sample similarity is high in the process of image classification for deep learning network,based on AlexNet,the convolution neural network model,we present the adoption of data preprocessing methods like data set amplification,background segmentation and principal component analysis for data set. The basic structure of the convolution neural network model is five convolution layers,two full connection layers and dropout layers. The experimental environment is ubuntu16.04 system and Caffe,a depth learning framework. In the experiment,the original data set Leaves and the data set of apple surface lesion are tested in classification and recognition firstly,and their accuracy rate are respectively 84% and 78%. Then the data set after enhanced are tested again. The accuracy rate of Leaves is 86%,increased by 2%,and that of the apple surface lesion data set is 83%,improved by 5%. The test shows that the accuracy on the network of the data sets of leaves and apple surface lesion has been improved after data enhancement.

关 键 词：深度卷积神经网络 数据增强 识别精度 图像识别 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]