基于CIFAR-10的图像分类模型优化  

THE OPTIMIZATION OF IMAGE CATEGORIZATION MODEL BASED ON CIFAR-10

作　　者：张占军 彭艳兵[2] 程光 Zhang Zhanjun;Peng Yanbing;Cheng Guang(Wuhan Research Institute of Posts and Telecommunications,Wuhan 430074,Hubei,China;Nanjing Research and Development Department,FiberHome Telecommunication Technologies Co.,Ltd.,Nanjing 210019,Jiangsu,China)

机构地区：[1]武汉邮电科学研究院,湖北武汉430074 [2]烽火通信科技股份有限公司南京研发,江苏南京210019

出　　处：《计算机应用与软件》2018年第3期177-181,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(61602114);国家高技术研究发展计划(2015AA015603)

摘　　要：随着卷积神经网络在图像处理的研究与应用,图像的分类准确度得到了大幅提升,但是过拟合的问题却一直存在并成为影响分类准确率的重要因素。从过拟合的产生源头出发,增加数据量并减少参数数量以达到降低过拟合的目的。基于经典模型Le Net-5,对输入数据进行数据增强,并对卷积层进行拆分以减少参数,同时采用L1、L2混合约束的方法,并灵活调整两者的占比以达到最佳效果。实验结果表明,在CIFAR-10数据集上,优化后的网络达到了91.2%的准确率,相比最初的Le Net-5模型提高了23%,极大地降低了过拟合,提高了模型的分类准确率。With the research and application of convolutional neural network in image processing,the accuracy of image classification has been greatly improved,but the problem of over-fitting has always existed and has become an important factor affecting the classification accuracy.In this paper,starting from the source of over-fitting,we increased the amount of data and reduced the number of parameters in order to reduce the over-fitting purposes.Based on the classical model LeNet-5,this paper made input data enhancement and split the convolution layer to reduce the parameters.At the same time,it used L1 and L2 mixed constraints and adjusted the proportion of the two to achieve the best effect.Experimental results showed that the optimized network achieved 91.2%accuracy on the CIFAR-10 dataset.Compared with the original LeNet-5 model,it was increased by 23%.It greatly reduced the over-fitting,and improved the classification accuracy of the model.

关 键 词：过拟合 数据增强 正则约束 卷积拆分 准确率 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]