作　　者： (杨涛）; (张云伟）; (杨瑞）;

机构地区： 昆明理工大学信息工程与自动化学院,云南昆明650504

出　　处： 《传感器与微系统》 2017年第8期59-61,68,共4页

摘　　要： 提出了一种基于深度信念网络(DBN)的社保卡号码识别方法,通过采集社保卡图像,采用模块分割的方法,对社保卡号码区域进行行分割,利用区域生长的方法对行内号码分割,将号码图像灰度化与二值化,并归一化为32×32大小,作为深度信念网络的输入数据,训练3层受限玻尔兹曼机(RBM)来获得更加抽象的特征表达,模型的最顶层结合Softmax回归分类器对抽取后的特征进行分类。实验结果表明:其准确率高达98.3%,与BP神经网络和支持向量机(SVM)模型相比,深度信念网络学习了数据的高层特征的同时降低了特征维数,提高了分类器的分类精度,有效提高了社保卡号码识别率。 A method based on deep belief networks（ DBN） is proposed to identify social security card number.Firstly,collect the social security card image and segment the card number area by the module segmentation.Secondly,split a single character of the card number using the regional growth method. Thirdly,the character image is grayed,binarized and normalized to the size of 32 × 32,which is taking as the input data of DBN.Training the 3 layers restricted Boltzmann machine（ RBM） to obtain more abstract features. The top layer of the model combined the softmax regression classifier to classify the extracted features. Experimental results show that the accuracy rate is up to 98. 3%. Compared with the BP neural network and support vector machine（ SVM）model,the DBN not only learned the high-level characteristics but also reduced the dimension of features and improved the accuracy of classifier classification. Finally,it effectively improved the recognition rate of social security card number.

关 键 词： 模块分割 深度信念网络 受限玻尔兹曼机 回归分类器