基于反馈知识迁移的脱机手写体汉字识别方法  

Approach for off-line handwritten Chinese characters recognition based on knowledge transfer with feedback

作　　者：朱程辉[1] 沈飞 王建平[1] 孙伟[1] ZHU Chenghui;SHEN Fei;WANG Jianping;SUN Wei(School of Electrical Engineering and Automation,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学电气与自动化工程学院,安徽合肥230009

出　　处：《传感器与微系统》2020年第5期30-33,40,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(51877060)。

摘　　要：针对卷积神经网络识别脱机手写体汉字时出现的梯度弥散问题,提出一种基于反馈知识迁移的识别方法。将卷积神经网络分解为主网络与若干子网络,使网络参数转移到低层数网络中。按字型结构分割脱机手写体汉字送入子网络中训练,再使用知识迁移将多个子网络的知识迁移到主网络中,结合反馈理论自适应调节子网络权重系数,实现对网络整体的知识整合。仿真实验表明:本文方法有效缓解了梯度弥散现象,具有较高的识别率。Aiming at the problems of gradient vanishing in recognition of off-line handwritten Chinese characters using convolutional neural network,a recognition approach based on knowledge transfer with feedback is given.Convolutional neural network is decomposed into a main network and several sub-networks,network parameters are transfered to networks with less layers.Chinese character samples are segmented and sent to sub-networks for training by type structure,then knowledge of several sub-network are transferred to main-network by knowledge transfer.Adjust the weight coefficient of sub-networks adaptively with feedback theory,to realize knowledge integration of the whole network.Simulation result shows that this approach relieves gradient vanishing effectively and has higher recognition rate.

关 键 词：脱机手写体汉字 梯度弥散现象 知识迁移 反馈 权重系数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]