改进BP神经网络在手写数字识别中的性能研究  

Research on Improving the Performance of BP Neural Network in Handwritten Digit Recognition

作　　者：张红 马静 Zhang Hong;Ma Jing(Wuhai Vocational and Technical College,Wuhai Inner Mongolia 016000,China)

机构地区：[1]乌海职业技术学院,内蒙古乌海016000

出　　处：《信息与电脑》2017年第19期56-57,60,共3页Information & Computer

基　　金：内蒙古自治区高等学校科学研究项目(项目编号:NJZC363)

摘　　要：为了正确反映经常采用的改进BP神经网络在手写数字识别中的性能,笔者选取常用的改进BP神经网络算法、动量BP算法、自适应学习速率BP算法、共轭梯度BP算法、LM-BP算法进行手写数字识别试验。笔者首先介绍BP神经网络的基本原理,指出其存在学习速率低、训练易陷入局部收敛的问题,并针对该问题提出学术界中的四种改进算法。介绍常用手写数字识别的预处理方法,包括图像二值化、图像平滑、字体的分割、字体细化、大小归一、去毛刺等。最后采用四种改进的BP神经网络算法对预处理后的手写数字样本进行训练和识别,并比较不同改进算法的性能。从实验结果来看,LM-BP算法能有效加快BP算法的训练速度,并具有较高的识别准确率。In order to correctly reflect the improved BP neural network on the recognition performance of handwritten numerals,the author selects the commonly used improved BP neural network algorithm-momentum BP algorithm,adaptive learning rate BP algorithm,conjugate gradient BP algorithm,LM-BP algorithm Handwritten digital recognition test.The author first introduces the basic principle of BP neural network,points out that there is a low rate of learning,training is easy to fall into the problem of local convergence,and puts forward four kinds of improvements in the academic field.This paper introduces the preprocessing methods of handwritten numeral recognition,including image binarization,image smoothing,font segmentation,font refinement,size normalization,deburring and so on.Finally,four improved BP neural network algorithms are used to train and identify the handwritten digital samples after pretreatment,and the performance of different improved algorithms is compared.From the experimental results,LM-BP algorithm can effectively speed up the training speed of BP algorithm,and has high recognition accuracy.

关 键 词：BP神经网络 性能比较 数字识别 特征提取 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]