半连续隐马尔科夫模型脱机阿拉伯手写识别  

Off-line Arabic Handwriting Recognition System Based on Semi-continuous Hidden Markov Models

作　　者：向东[1] 刘虎[1] 陈先桥[1] 程艳芬[1] 

机构地区：[1]武汉理工大学计算机科学与技术学院,湖北武汉430063

出　　处：《武汉理工大学学报（信息与管理工程版）》2011年第3期349-353,共5页Journal of Wuhan University of Technology：Information & Management Engineering

基　　金：武汉理工大学自主创新研究基金资助项目

摘　　要：提出一种基于半连续隐马尔科夫模型的脱机阿拉伯手写识别系统。该系统采用滑动窗口技术直接从数字图像提取像素密度和凹度特征,没有复杂的预处理操作,每个窗口被划分为前景像素数目相同的4个子窗口,前后相邻窗口的重叠为窗口宽度的1/2,且特征的维数仅为36维,节约了系统的计算量。采用半连续隐马尔科夫模型对字符进行建模,而词的模型采用嵌入式训练方法,无需对词进行预先分割,从而避免了切分引入的误差。在IFN/ENIT测试数据库上的测试结果表明,一个最佳输出时平均识别率能够达到86.6%。An off-line Arabic handwriting recognition system based on semi-continuous hidden Markov models was proposed.The system extracts pixels density and concavity features directly on binary images without complex preprocessing.There are same foreground pixels in the sub-block of each window.The overlap between successive windows is half width of window and the dimension of feature is only 36,these would save the computation of the system.The proposed system builds character models and learns word models using embedded training without character pre-segmentation.Experiments implemented on the benchmark IFN/ENIT database show that the average recognition rate of this system with one best output is 86.6%.

关 键 词：文字识别 脱机 阿拉伯手写 半连续隐马尔科夫模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]