基于特征图叠加的脱机手写体汉字识别  

Offline Handwritten Chinese Character Recognition Based on Concatenated Feature Maps

作　　者：毛晓波[1] 程志远 周晓东 MAO Xiaobo;CHENG Zhiyuan;ZHOU Xiaodong(School of Electrical Engineering,Zhengzhou University,Zhengzhou 450001,China;Industrial Technology Research Institute,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]郑州大学电气工程学院,河南郑州450001 [2]郑州大学产业技术研究院,河南郑州450001

出　　处：《郑州大学学报（理学版）》2018年第3期78-82,共5页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(61473266);河南省高校科技创新团队支持计划项目(17IRTSTHN013);河南省高等学校重点科研项目(17A413011)

摘　　要：采用深度学习方法处理脱机手写体汉字识别问题,提出一种改进的卷积神经网络结构.不同于传统的层与层依次连接的方式,新的结构中将当前层与前一层的特征图进行叠加后共同作为输入进入下一层.与以往结构相比,该卷积神经网络结构缓解了梯度消失的问题,减少了参数数量,且更有效地保留了汉字的细微结构特征.在ICDAR-2013脱机手写体汉字识别竞赛测试集上的实验结果表明,该结构可以在减少模型参数的情况下得到较高的识别准确率.An improved architecture of convolutional neural network was proposed by using deep learning method to deal with offline handwritten Chinese character recognition.Different from the traditional way connected layer by layer,the new architecture concatenated the feature maps from the current layer and the previous layer as common inputs to the next layer.The new neural network architecture alleviated the vanishing-gradient problem,reduced the number of parameters and preserved the sub-structure of Chinese characters more effectively.The results on the ICDAR-2013 offline handwritten Chinese character recognition competition test set showed that the proposed method had the higher recognition accuracy rate in the case of reducing parameters.

关 键 词：汉字识别 深度学习 卷积神经网络 特征图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]