Computer Engineering and Applications ›› 2017, Vol. 53 ›› Issue (20): 187-194.DOI: 10.3778/j.issn.1002-8331.1701-0305

Previous Articles     Next Articles

Weighted fusion of texture features based Central Asian multi-scripts identification

Buvajar Mijit1, Kurban Ubul1, Nurbiya Yadikar1, Tuergen Yibulayin1, Alimjan Aysa2   

  1. 1.Institute of Information Science and Engineering, Xinjiang University, Urumqi 830046, China
    2.Network and Information Center, Xinjiang University, Urumqi 830046, China
  • Online:2017-10-15 Published:2017-10-31



  1. 1.新疆大学 信息科学与工程学院,乌鲁木齐 830046
    2.新疆大学 网络与信息中心,乌鲁木齐 830046

Abstract: Many similar shaped scripts are used all over the world today, script identification with similar shaped characters is difficult task in pattern recognition area, and it is one of the urgently solved problems. However, there are a few reports for identification of Central Asian countries and Chinese Minority scripts, especially for scripts with similar shaped characters. In this paper, firstly, two multi-script document image databases are established, which are including 1, 600 and 2, 200 plain text document images respectively in 11 scripts such as English, Chinese, Russian, Mongol, Arabic, Tibet, Uyghur, Turkish, Uzbekistan, Tajikistan and Kazakhstan. Then, six texture features such as mean, standard deviation, entropy, consistency, third order moment and smoothness are extracted from whole page image respectively, and they are classified using seven different kinds of classifiers. On the basis of finding the sensitivity of each feature for the document image, it is determined the optimal weights suitable for identification of central Asian multilingual scripts after the weighted fusion method is used to extract the fusion features. Finally, they are classified by using different classifier via multi- features weighted coefficient fusion, and it is obtained 99.38% and 95.42% of average identification rate with the two established dataset separately. Experimental results indicate that texture features and weighted fusion texture features can better describe the multi-script document images, and they can effectively classify these 11 kinds scripts mentioned above.

Key words: script identification, texture feature, discriminant analysis, Mahalanobis distance, weighted fusion

摘要: 全球各地目前使用很多种相似的文种,相似文种的识别是模式识别领域内难点并迫切需要解决的问题之一。然而,针对中亚文种文本文档和少数民族文种也就是相似文种分类识别方面的文献报道几乎没有。首先建立了两个多文种文档图像数据库,分别有1 600幅和2 200幅纯文本整篇文档图像,包含英文,汉文,俄文,蒙文,阿拉伯文,藏文,维吾尔文,土耳其文,乌兹别克文,塔吉克文和哈萨克文等共有11种文档图像。其次分别提取文档图像的均值,标准差,熵,一致性,三阶矩,平滑度等六个纹理特征,利用不同7种分类器分类。在找到各个特征对多文种文本文档图像的灵敏度的基础上,采用加权特征融合方法提取融合特征,确定了适合中亚多文种文档图像识别的最佳权值。最后用不同分类器分类识别,通过多特征以系数加权融合之后,以建立的两个数据库基础下获得平均的识别率分别为99.38%和95.69%。实验结果表明,提取的纹理特征和加权融合的纹理特征能较好地描述文档图像特征,并且它们可以有效地分类以上所述的11个文种。

关键词: 文种识别, 纹理特征, 判别分析, 马氏距离, 加权融合