【作者】 杨基春；

【导师】 黄战华；

【作者基本信息】 天津大学 ， 光学工程， 2010， 博士

【摘要】 轮胎规格号是指生产厂家在轮胎胎侧印制的代表轮胎规格的特定意义的字符,对于厂家来说,这些参数在轮胎工厂管理中对于轮胎类型的识别分类非常重要,可用来作为生产管理、质量追查的一个重要参量。目前OCR技术被广泛用于字符识别,但在国内还没有轮胎规格号识别的相关研究及报导,国外针对DOT(US Department of Transportation缩写)码和轮胎规格号识别的一个比较成熟的技术是线激光扫描CCD成像识别,这套识别系统采用三角法测量技术,优点是识别轮胎模压成型字符可靠,强健,缺点是成本高,处理时间长,需要线激光发射设备、CCD成像系统和机械运动结构,并且对印刷字符无法识别。本文在分析国内外研究现状的基础上,围绕轮胎规格号识别开展研究,针对目标图像的特点,提出LED阵列照明、CMOS成像以及BP神经网络算法识别系统,该系统成本低,速度快,能有效识别不同字体、大噪声背景下的轮胎规格号字符。其中,在图像预处理方面,采用最小二乘法迭代拟合圆法对圆心进行精确定位,并且提出利用积分加窗统计法提取规格号字符区域,用边缘自适应定位算法对单个字符进行分割,有效排除了大噪声干扰。在规格号识别上,提取了多种特征分析,利用BP神经网络算法的字符分类器和基于规格号关系特征的二次分类器识别,使得规格号识别达到了较高的识别率,并且该识别系统耗费时间少,能够满足快速检测要求,本课题的研究工作取得了预期效果。本文工作的主要创新点如下:1、轮胎图像采集的照明技术研究。由于轮胎图像背景和前景字符是同种橡胶材质,表面色差很小,因此通常采集得到的图像有噪声大、对比度低的特点,本文采用特定设计的环形阵列LED结构照明,有效降低了图像噪声,提高图像对比度。2、高噪声背景下的识别目标提取算法。算法利用积分窗口来降低图像边缘图维数,由二维变成一维,并提出利用一维信号的极值点来形成锯齿波,依据锯齿波的波形参数和其规则性来提取规格号字符区域;针对字符个体分割的复杂性问题,提出边缘自适应定位算法,该算法利用边缘强度、连接强度和字符宽度,采用阈值由小到大、区域分割由大到小多次迭代来定位字符。算法抗噪能力强,适合复杂环境下的字符精确定位。3、结合关系特征的轮胎规格号综合识别算法。针对规格号字符特点,深入研究了轮胎规格号字符识别特征,提出了基于BP神经网络的规格号字符识别;并提出基于规格号关系特征的二次分类器识别算法;针对噪声混淆字符信息的情况,提出字符截断方法以获得不完全可区分字符特征算法,提高字符识别率。实验结果表明,结合关系特征的轮胎规格号综合识别算法适合于轮胎规格号识别。更多还原

【Abstract】 Tire Specifications is character printed in sidewall of the body by manufacturers on behalf of specific significance. These parameters is very important to identify the type of classification for tire factory management, used as a important parameters for production management、quality of tracing.At present, OCR technology is widely used in character recognition, but tire specifications have not yet identified its research and related reporting in the domestic. A more mature technology abroad for the DOT Code and Tire Specifications identification is the CCD imaging with scanning of line laser, this identification system using triangulation measurement technology, the advantage of identification of tire molding character are reliable and robust, the disadvantage are high cost、long processing time and needing line laser equipment、CCD imaging systems and mechanical structure, moreover, it does not recognize the printed characters.Based on the analysis of the history and recent developments around tire Specifications identification, study was carried out. The article analysed characteristics of the target image, a system was presented that included LED lighting arrays, CMOS imaging, as well as identification of BP neural network, which low-cost, fast and effective identification of different fonts of tire specifications with big background noise. Among them, in image pre-processing, the use of iterative least square method used to fit accurate positioning of the circle center , and proposed integral windowed statistics to extract specifications character region, a edge adaptive localization algorithm for segmentation of individual characters,which effectively ruled out big noises. In specifications recognition, research was developed by using a variety of characteristics extracted, and combined methods were used to identify tire specifications with classifier of BP neural network algorithm-based classifier and secondary classifier recognition algorithm of specifications, which have high recognition rate and the less time-consuming. The identification system meeted the rapid testing requirements and came up to the desired effect.The following are the main innovations.1. Lighting technology research for Tire image capture.As tire image background and character prospects is the same kinds of rubber material, the surface color difference is very small, therefore the image acquisition are with noise and low contrast, this paper adopts specific design structure of a ring array of LED lighting, effectively reducing the image noise, increase image contrast.2. Extraction algorithms for recognition object of high noise background. Algorithm using integral window to reduce the dimensions of the image edge map, from two-dimensional to one-dimensional, and to make use of one-dimensional signals to form the sawtooth extreme points, based on sawtooth waveform parameters and specifications of its rules to extract the number of characters in the region ; for the segmentation of individual characters in the complexity of the issue and put forward edge adaptive localization algorithm, the algorithm uses edge-connection strength、edge strength and character length and using threshold from small to large, regional division multiple iterations to locate the characters in descending . Algorithm for anti-noise capability, suitable for a complex environment, precise positioning of characters.3.Combining with the relation features of an integrated recognition algorithm of tire specifications. According to the characteristics of specification characters, in-depth study of the tire specification character recognition features, recognition algorithm of specifications was presented based on BP neural network character recognition; and secondary classifier recognition algorithm of specifications was proposed based on the relationship features . For the noise, confusion of character information proposed method to obtain the truncated characters are not fully distinguish between characters feature algorithm to improve character recognition rate. Experimental results show that the comprehensive specifications recognition algorithm combined with relationship features is adaptive of tire specifications recognition..更多还原