自然场景下字符区域的定位与识别

发布时间：2018-09-17 09:48

【摘要】：随着手机等各种智能移动设备的发展,人们越来越关注对生活中及工业环境等自然场景下文本信息的获取与利用。自然场景中的文本信息不同于印刷文本,它具有多样的语言文字、不同的字体和大小、复杂背景的干扰、残缺和污损的影响,这些都为自然场景下文本信息的获取与利用带来一定的困难和挑战。本论文致力于对自然场景下文本区域准确定位和识别的研究,重点研究英文字符与数字区域的定位与识别;并在此基础上进一步研究自然场景文本区域定位与识别的一种特殊场景,复杂工业环境下的铁路油罐车车号区域定位与识别。铁路油罐车车号区域作为自然场景下文本区域的一部分,因其字符存在断裂等特点,本论文把复杂工业环境下的铁路油罐车车号区域定位与识别作为自然场景下文本区域定位与识别的一种特殊场景进行研究,以达到从复杂工业环境下的众多干扰中准确定位出铁路油罐车车号区域的目的,并对其进行断裂字符的分隔与识别。本论文在充分对比总结各种文本区域定位方法的基础之上,给出一种既适用于自然场景下普遍存在的英文字母和数字区域的定位,也适用于复杂工业环境下铁路油罐车车号区域定位的通用方法。该方法对尺寸大小不同、存在倾斜、受光照变化影响的文本区域具有较好的定位效果,首先利用最大稳定极值区域(Maximally Stable Extremal Regions,MSER)检测获取极值区域,并从得到的极值区域中筛选、连接成有效区域对,进一步由有效区域对得到三联体区域,然后由符合一定特征的三联体区域连接成候选文本区域,并利用支持向量机(Support Vector Machine,SVM)对候选文本区域进行筛选。该方法对自然场景下的文本区域具有较好的定位效果,为验证本方法的通用性,本论文对自然场景下一个特殊应用场景-复杂工业环境下铁路油罐车车号信息的获取进行研究。本论文充分针对铁路油罐车车号字符断裂的特点,给出一种适合于断裂字符的分隔方法。对于字符的识别考虑到自然场景下的字符种类繁多、字体多变,而铁路油罐车车号字符种类固定、字体变化相对较少,采用不同的方法对这两大应用场景中的字符进行识别。对于自然场景下普遍存在的英文字母和数字,使用Tesseract-OCR进行训练与识别;对于铁路油罐车车号字符,本论文利用SVM对其进行分类识别。经过大量实验证明,本论文所采用方法对两大研究场景中的文本信息具有较好的定位与识别效果。
[Abstract]:With the development of various smart mobile devices such as mobile phones, people pay more and more attention to the acquisition and utilization of text information in natural scenes such as life and industrial environment. Text information in natural scenes is different from printed text, it has a variety of languages, different fonts and sizes, complex background interference, damage and the impact of defacement, All these bring some difficulties and challenges to the acquisition and utilization of text information in natural scenes. This thesis is devoted to the research of the exact location and recognition of the following regions of the natural scene, focusing on the localization and recognition of the English character and digital regions; On this basis, a special scene of text location and recognition of natural scene, regional location and recognition of railway tanker number in complex industrial environment is further studied. As a part of the following area of natural scene, the railway tanker car number area is characterized by the rupture of its characters. In this paper, the location and recognition of the railway tanker number in complex industrial environment is studied as a special scene of the location and recognition of the local area in the following natural scene. In order to locate the area of railway tanker number accurately from many disturbances in complex industrial environment, and to separate and recognize the broken characters. On the basis of comparing and summarizing all kinds of text region localization methods, this paper presents a new location method which is suitable for English alphabet and digital region, which is widely used in natural scene. It is also applicable to the regional location of railway tanker number in complex industrial environment. This method has better localization effect for text regions with different sizes and tilting, which are affected by illumination changes. Firstly, the maximum stable extremum region (Maximally Stable Extremal Regions,MSER) is used to detect and obtain the extremum region, and the obtained extreme value region is screened. Then the triplet region is connected to the candidate text region by the effective region pair, and then the candidate text region is selected by using support vector machine (Support Vector Machine,SVM). In order to verify the generality of this method, this paper studies the acquisition of railway tanker number information in complex industrial environment, which is a special application scene in natural scene. In this paper, according to the character of railway tanker car number character break, a separation method suitable for breaking character is given. The recognition of characters takes into account the variety of characters and the variety of fonts in natural scenes, while the type of characters of railway tanker car number is fixed and the font changes are relatively little, so different methods are adopted to recognize the characters in these two major application scenarios. Tesseract-OCR is used to train and recognize the common English letters and numbers in natural scenes, and SVM is used to classify and recognize the vehicle number characters. A large number of experiments have proved that the method adopted in this paper has a good effect on the location and recognition of text information in the two research scenarios.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】