面向手写体识别的DCNN抗噪性能研究  

Research on Anti-noise Performance of DCNN for Handwriting Recognition

作　　者：蔡际杰 陈德旺 张璜 CAI Ji-jie;CHEN De-wang;ZHANG Huang(College of Mathematics and Computer Science,Fuzhou University;Key Laboratory of Intelligent Metro of Universities in Fujian Province,Fuzhou University,Fuzhou 350108,China;School of Computing and Information Science,Fuzhou Institute of Technology,Fuzhou 350506,China)

机构地区：[1]福州大学数学与计算机科学学院 [2]福州大学智慧地铁福建省高校重点实验室,福建福州350108 [3]福州理工学院计算与信息科学学院,福建福州350506

出　　处：《软件导刊》2021年第5期1-6,共6页Software Guide

基　　金：国家自然科学基金项目(61976055);智慧地铁福建省高校重点实验室建设基金项目(53001703,50013203)。

摘　　要：深度卷积神经网络(DCNN)是人工智能研究领域前沿方向。DCNN结构复杂、参数非常多、可解释性与鲁棒性不强,对图像数据集的清晰度要求很高,而目前关于DCNN抗噪性能研究还较欠缺。通过给手写体数据集DigitDataset的测试集添加4种不同幅度噪声,深入研究DCNN在手写体识别上的抗噪性能。研究结果表明:①噪声对DCNN性能影响很大,噪声幅度越大,精度下降越快;②指数噪声对精度影响最大,伽马噪声、瑞利噪声次之,高斯白噪声影响最小;③随着噪声参数a和参数b的增大,识别精度大幅度下降。该结果对DCNN的改进和高鲁棒性的深度学习系统(如深度模糊系统等)研究具有一定参考价值。Deep convolutional neural network(DCNN)is a popular and cutting-edge direction in the field of artificial intelligence research,which is widely used in various fields.However,due to the complex structure,many parameters,poor interpretability and robustness of DCNN,it has high requirements for the clarity of image data sets.At present,there are some reports about the low antinoise performance of DCNN,but in-depth research on its anti-noise performance is still lacking.By adding four kinds of noises with different amplitudes to the test set of the handwriting data set DigitDataset,the anti-noise performance of DCNN on handwriting recognition problems is deeply studied.The research results show that:①Noise has a great impact on the performance of DCNN.The greater the noise amplitude,the faster the accuracy decline;②Exponential noise has the greatest impact on its accuracy,followed by gamma noise and rayleigh noise,and gaussian white noise the impact is minimal;③With the increase of noise parameter a and parameter b,the recognition accuracy is greatly reduced.The research results have certain reference significance for the improvement of DCNN and the research of highly robust deep learning systems(such as deep fuzzy system,etc).

关 键 词：深度卷积神经网络 抗噪性能 手写体识别 鲁棒性 可解释性 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]