基于离散曲波变换的多文种文档图像文种识别  

Script identification of multi-script document images based on discrete curvelet transform

作　　者：李顺 木特力铺.马木提 吾尔尼沙.买买提 阿力木江.艾沙[2] 库尔班.吾布力 LI Shun;Mutelep·Mamut;Hornisa·Mamat;Alim·Aysa;Kurban·Ubul(School of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Network and Information Center,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046 [2]新疆大学网络与信息中心,新疆乌鲁木齐830046

出　　处：《计算机工程与设计》2019年第5期1376-1382,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61563052;61363064;61862061);新疆大学博士科研启动基金项目(BS150262);新疆维吾尔自治区高校科研计划创新团队基金项目(XJEDU2017T002)

摘　　要：为提高文种识别效果,提出一种基于离散曲波变换的文种识别方法。利用文档图像经过曲波变换后得到的cell矩阵中的实数曲波系数,提取共82维能量特征;使用贝叶斯、KNN和判别分析3种分类器进行训练和分类。对两个数据库进行实验,数据库1包含8种文字共1600幅图片,使用3种分类器得到平均大于99%的识别准确率;数据库2包含10种文字共10 000幅图片,得到平均大于98%的识别准确率。实验结果表明,该方法运算速度快,具有良好的鲁棒性,识别效果优于基于小波变换的文种识别方法和基于二元复数小波变换的文种识别方法。To increase the effect of script identification,a method of script identification based on discrete curvelet transform was proposed.The real curvelet coefficient of the cell matrix was obtained from the document image transformed by curvelet transform,and totally 82 dimensional energy features were extracted.Bayesian,KNN and discriminant analysis were used to train and classify,and two databases were used for experimentation.The experimental database I,which contained 8 scripts and 1600 images,was classified by the three classifiers.The average accuracy rate obtained is more than 99%.The experiment database II contained 10 scripts and 10 000 images,which were classified by the three classifiers.An average accuracy rate obtained is more than 98%.Experimental results show that the proposed method is fast with good robustness.The recognition effect is better than that based on wavelet transform and that based on dual-tree complex wavelet transform.

关 键 词：文种识别 离散曲波变换 实数曲波系数 判别分析 鲁棒性 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]