统计学全局特征提取在纹理图像光学字体识别中的应用

The application of global feature extraction with statistical analysis in OFR from texture images

针对纹理图像光学字体识别中大部分现有方法容易出现噪声干扰和高度细分依赖的问题,提出了一种基于统计学分析的全局特征提取方法。首先,使用二值化、倾斜校正和文本归一化预处理图像,得到完整的文本块图像;然后,使用拉普拉斯过滤器降低图像的椒盐噪声,并利用反相过滤器得到白色背景和黑色边缘的图像;最后,基于模式边界边缘像素之间关系进行统计学分析以提取出22个有用特征,借助分类器完成识别。在阿拉伯文书法脚本图像数据集上的实验结果表明,本文方法使用决策树分类器可获得最好的分类精度,高达98.26%,相比其他的较为新颖的特征提取方法,本文方法获得更好的识别性能。

For the issue that a majority of existing approaches have noise and high subdivision dependency in optical font recognition from texture image,a global feature extraction approach based on statistical analysis is proposed.Firstly,binary,tilt correction and text normalization is used to preprocessing images to generate total images with text block. Then,Laplace filter is used to reduce impulse noise and reverse filter is used to generate images with white background and black edge. Finally,statistical analysis is done based on the relationship between pattern edge pixels to extract 22 useful features,and classifier is used to finish recognition. Experimental results show that proposed method can get the best classification accuracy with 98. 26% when using decision tree classifier. And it has better recognition performance than other advanced feature extraction method.