基于多流卷积神经网络的中文笔迹鉴别研究  

Multi-stream convolutional neural network for offline Chinese writer identification

作　　者：毛颖裕 张怡龙 王海霞 MAO Yingyu;ZHANG Yilong;WANG Haixia(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023;College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]浙江工业大学信息工程学院,杭州310023

出　　处：《高技术通讯》2023年第8期849-859,共11页Chinese High Technology Letters

基　　金：国家自然科学基金(61976189,61905218)资助项目。

摘　　要：针对中文笔迹签别中局部特征与全局特征的提取与融合问题,本文提出一种基于多流卷积神经网络的中文离线笔迹签别方法。该网络共有3条支流结构,其中1条支流对整图进行全局特征提取,另外2条支流分别对原图采用水平分割和垂直分割方式来获取局部数据并进行特征提取,而后将提取到的特征与全局特征进行多尺度融合。本文采用了top-1和top-5 2种准确率进行性能评估并进行了作者级别的准确率对比分析。实验结果表明,本文所提出的结构不仅可以更好融合、利用局部和全局特征还能拓宽网络获取细节特征的来源,改善网络性能。Aiming at the extraction and fusion of local and global features in Chinese handwriting identification,this paper proposes a multi-stream convolutional neural network based offline Chinese writer identification method.The network has three tributary structures,one of which extracts global features from the entire image while the other two perform feature extraction on image segments cut horizontally and vertically from the original image,respectively.Multi-scale fusion is then performed on the extracted features.This paper uses top-1 and top-5 accuracy rates for performance evaluation and carries out the comparative analysis based on accuracy rate of each author.The experimental results show that the proposed method can not only integrate local and global features,but also broaden the source of the network to obtain detailed features,consequently improving the network performance.

关 键 词：笔迹鉴别 特征提取 模式识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]