• 首页
  • 期刊简介
  • 编委会
  • 投稿指南
  • 收录情况
  • 杂志订阅
  • 联系我们
引用本文:谢璐阳,夏兆君,朱少华,张代庆,赵奉奎.基于卷积神经网络的图像识别过拟合问题分析与研究[J].软件工程,2019,22(10):27-29.【点击复制】
【打印本页】   【下载PDF全文】   【查看/发表评论】  【下载PDF阅读器】  
←前一篇|后一篇→ 过刊浏览
分享到: 微信 更多
基于卷积神经网络的图像识别过拟合问题分析与研究
谢璐阳,夏兆君,朱少华,张代庆,赵奉奎
(南京林业大学汽车与交通工程学院,江苏 南京 210037)
摘 要: 近年来深度学习在很多领域发挥着重要作用,但是在训练过程中存在模型过拟合的问题。针对该问题, 本文对Kaggle竞赛中典型的猫狗识别任务建立了卷积神经网络,并分析研究了多种抑制过拟合的方法,包括添加L2正 则项、dropout处理、数据增强及多种方法综合使用的综合法,分别分析不同方法在训练集和验证集上的训练精度和 损失,发现数据增强法优于其他两种方法,且综合法能够消除过拟合。研究结果对卷积神经网络的配置具有重要的参 考价值。
关键词: 卷积神经网络;过拟合;图像识别;深度学习
中图分类号: TP311    文献标识码: A
基金项目: 江苏省高等学校自然科学研究面上项目“X射线荧光光谱信号分析理论研究”(项目编号:17KJB150024).
Analysis and Research of Overfitting of Image Recognition Based on Convolutional Neural Networks
XIE Luyang,XIA Zhaojun,ZHU Shaohua,ZHANG Daiqing,ZHAO Fengkui
( College of Automobile and Traffic Engineering, Nanjing Forestry University, Nanjing 210037, China)
Abstract: Deep learning is playing important roles in various fields.However,it suffers overfitting in the training process.Dog and cat recognition is a classical task in Kaggle competition.Based on this problem,a convolutional neural network is created to analyze the effects of different methods for correcting overfitting problems.Those methods include L2 regularization,dropout,data augmentation and comprehensive method.Accuracy and loss of the model from training and validation sets are used to analyze the performance of those measures.Results show that the method of data augmentation performs better than L2 regularization and dropout,and the best method is the comprehensive method which eliminates overfitting in this case.The result is of great significance for configuring Convolutional Neural Networks.
Keywords: Convolutional Neural Networks;overfitting;image recognition;deep learning


版权所有:软件工程杂志社
地址:辽宁省沈阳市浑南区新秀街2号 邮政编码:110179
电话:0411-84767887 传真:0411-84835089 Email:semagazine@neusoft.edu.cn
备案号:辽ICP备17007376号-1
技术支持:北京勤云科技发展有限公司

用微信扫一扫

用微信扫一扫