低质量印刷体字符分割与识别研究

【作者】 孙强；

【导师】 娄震；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2014， 硕士

【摘要】 印刷体字符识别技术发展的越来越成熟,在现代社会中的应用越来越广,但是低质量印刷体字符的识别依然是印刷体字符识别技术发展的瓶颈。本文分析了低质量印刷体字符识别过程中常遇见的问题,根据这些常出现的问题结合本文实验对象的特点和规律,设计了一个低质量印刷体字符识别流程,并通过车牌字符区域图像、人民币冠字号码区域图像作为实验对象来验证本文介绍的识别过程的有效性。二值化阈值选取算法中,本文使用了在Otsu算法获得的全局阈值的基础上,通过选取该阈值邻域内的值作为候选阈值,结合连续字符连通域分析评价准则得出的评价分数,从候选阈值出选出最优阈值。该方法主要用在字符分割中,从实验效果图和分割实验结果显示了该方法的可行性。使用单一的字符分割策略对于低质量印刷体字符的分割能力有限,在字符分割阶段,本文采用了基于字符轮廓连通域分析与投影切分相结合的两级字符分割作为字符切割的方法。通过连续若干个印刷体字符连通域分析评价准则得出的评价分数,将两种分割方法串联起来,当一级字符分割评价分数低于某个阈值时,进入二级字符分割。实验表明,单使用一级字符分割方法没有两级字符分割方法分割正确率高。低质量印刷体字符图像中存在的干扰因素,使得相似字符的识别正确率往往要比其他字符的识别正确率低。为了减少相似字符之间的混淆错误,在字符识别阶段,本文使用基于敏感区域训练的二级字符识别来区分容易出现混淆错误的相似字符对,并利用遗传算法来获取字符识别阶段所需的一级可信度阈值参数和二级可信度阈值参数。实验表明,该方法能够有效的减少相似字符对之间的混淆错误。2021张车牌字符图像经过本文介绍的方法识别正确率达到了82.1%,比使用单一策略识别方法提高了3.3%；5113张人民币冠字号码图像识别正确率达到了93.7%,比使用单一策略识别方法提高了1.5%。更多还原

【Abstract】 The technology of machine printed characters recognition has been becoming more and more mature. More and more applications of machine printed characters sprung up. However, the recognition of the low quality of printed characters is still the bottleneck of OCR. This paper analyzes the problems which will happen when recognizing the low quality characters. In this paper a printed character recognition process is designed, the process considers the problems of low quality characters and the features of the experimental subject. In this paper the license plate images and the pictures of the crown word number area are used to validate the process.This paper uses the global threshold which is calculated by the Otsu algorithm to decide some candidate of thresholds. The best threshold is chosen from the candidates by the evaluation score which is computed by the evaluation criterion of the connected domain analysis of the continuous adjacent characters. This method of choosing the optional threshold is used in the stage of the character segmentation. The results of the character segmentation validate this method.The effect of using a single character segmentation strategy for the low quality printed characters is limited. In the stage of the character segmentation, a two level character segmentation strategy is used. The connected domain analysis of the continuous adjacent characters is used in the first level of the segmentation. If the calculated evaluation score of the analysis is greater than the pre-defined threshold, the second level character segmentation based on the analysis of the projection of the character image is used. The results of the character segmentation indicate the result of using two level character segmentation strategies is better than a single character segmentation strategy.There exit many interference factors in the low quality printed character pictures. This leads to the recognition correctness of the similar characters is less than the common characters. To reduce the error of the similar characters, a two-stage recognition method based on the sensitive area of the similar characters is introduced. The confidence coefficients are calculated by the genetic algorithm. The experiments demonstrate that this method can reduce the errors of the confusions which are caused by the similar characters.Using the method introduced in this paper the recognition correctness of2021car license plates reaches82.1%, which has the3.3%improvement compared to the single strategy. The correctness of5113crown word number pictures reaches93.7%, which has the1.5%improvement compared with the single recognition strategy.更多还原