基于VGG网络的古籍版面图像差异性比较方法  

The method of historical document image difference comparison based on VGG network

作　　者：翟立波 田学东 ZHAI Libo;TIAN Xuedong(School of Computer and Cyber Science and Technology,Hebei University,Baoding Hebei 071000,China)

机构地区：[1]河北大学网络空间安全与计算机学院,河北保定071000

出　　处：《激光杂志》2020年第9期121-126,共6页Laser Journal

基　　金：国家科学自然基金项目(No.61375075);河北省自然科学基金项目(No.F2019201329);河北省教育厅河北省高等学校科学技术研究重点项目(No.ZD2017208)。

摘　　要：针对传统特征提取方法在对古籍文字图像进行匹配时准确率较低,从而影响古籍版面图像差异性比较准确性的问题,提出基于VGG(Visual Geometry Group)网络的古籍版面图像差异性比较方法。该方法首先由VGG网络模型利用构建的差异性文字图像库进行训练得到卷积神经网络分类器;其次,对版面图像进行文字切分得到单字图像,将其输入到分类器中获取有效的字符级比对结果;最后,利用该结果,并对滑动窗口比较算法加以改进,利用键值对形式的数据结构来存储比较位置和比对结果,通过对比较位置进行映射的方式,避免比较过程中重复位置文字的二次比对问题,最终得到待比对的2幅古籍版面图像的文字差异位置并予以标记,以便于文献版本学研究。在对不同阁本《四库全书》版面图像的实验中,其标记准确率为89.7%,表明该方法有效提高了版面图像差异性比较的准确性。In view of traditional extraction way reveals a low accuracy when it comes to match ancient Chinese character images,exerting an effect on the accuracy of antique document image difference comparison,this paper,therefore,proposes a difference comparison method based on VGG(Visual Geometry Group)network for the image.In the method,the first step is to obtain convolution neural network classifier by training the VGG network model with the constructed differential character image library.Secondly,the whole image is segmented into a single character image and to be input into the classifier for the effective comparison results in character level.Finally,based on the result above to improve the sliding window comparison algorithm,the positions and results are compared via data structure in the form of key value pair.The mapping is carried out through the positions comparison to avoid the second comparison of repeated position character during the process.The character difference position of the two document images to be compared is attained and marked for the science of editions of Chinese ancient book study.In the experiment of document images in different versions of《SikuQuanshu》with the method,for instance,the mark accuracy is 89.7%,showing the accuracy of the difference comparison of the document images can be effectively bettered.

关 键 词：古籍版面 文字图像分类 差异性比较 VGG网络模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]