嵌入式字符识别技术的研究与开发

发布时间：2018-03-21 15:24

本文选题：字符识别　切入点：随机圆检测　出处：《江南大学》2012年硕士论文　论文类型：学位论文

【摘要】：随着工业生产的自动化程度越来越高,许多场合下需要对生产线上产品的标号或编码通过机器进行自动识别。然而工业现场的复杂环境,如光照不均匀、噪声等因素,给字符的准确识别带来较大的困难,对字符识别系统的硬件和软件的性能也提出了更高的要求。本文研究基于嵌入式智能相机硬件平台,以轴承压印字符为识别对象,在总结前人经验和成果基础上,提出了自己的解决方法,实现轴承压印字符的自动识别。论文所涉及的内容包括以下几个方面: 1)轴承图像的采集和预处理。在分析轴承压印字符的成像特点的基础上,设计了适用于轴承的图像采集系统。该系统基于嵌入式智能相机图像采集平台,并选用低角度LCD环形光源作为辅助光源,通过对轴承进行图像采集证实该系统能够获取质量较高的轴承图片。在图像预处理阶段,由于轴承是圆形结构且字符区域在图像中的位置不固定,因此需要对轴承字符区域进行定位和矫正,以便后续的处理和识别。为此,在使用大津法得到轴承的二值图像后,首先要对轴承圆心进行定位。针对随机Hough变换圆检测算法的不足,使用一种改进的随机圆检测算法用于轴承的圆心定位,在候选圆选取的过程中通过简单的判断降低了运算量,实验证实该圆心定位算法能够快速而准确定位出轴承圆心。随后,以定位出的圆心为极坐标原点,采用投影法确定轴承字符区域。最后,对轴承字符扇形区域进行矫正,若采用极坐标变换所得图像毛刺较多,因此本文采用基于仿射变换的字符矫正方法,实验表明该方法能使变换后的图像毛刺较少且速度上也能满足要求。 2)轴承字符的特征提取和分类器。本文详细描述了两种常用的特征提取方法:方向线素特征和轮廓层次特征,并在理论分析和实验的基础上,针对他们的不足分别提出了基于弹性网格的方向线素特征和基于小波变换的轮廓层次特征两种改进的特征提取方法。基于弹性网格的方向线素特征有效地弥补了均匀网格划分对字符形变的敏感性,特征的鲁棒性更强。基于小波变换的轮廓层次特征充分利用了小波多分辨率分析的特性,特征抗噪性更好且维数更低。关于轴承字符识别的分类器,本文讨论了BP神经网络和支持向量机分类器的原理,并分别设计了基于动量项与自适应学习速率的BP神经网络和基于LIBSVM的支持向量机网络。轴承字符中识别系统具体采用的特征提取方法和分类器方案是通过最后的识别实验确定的。通过对实际拍摄的图片进行识别实验,验证了本文所述方法的有效性,并确定特征提取方法和分类器的选取。实验表明:本文采用的图像采集方案能够获取质量较高的轴承图像,预处理方法效果较好,根据试验结果最终选取改进的轮廓层次特征作为特征,使用支持向量机作为分类器。字符识别的准确率在96%以上,速度上也能满足实际需求。本文方法能够快速而准确地对轴承压印字符进行自动识别。
[Abstract]:With the industrial production of increasingly high degree of automation, many applications need to automatically identify the product line or label encoding through the machine. However the complicatedindustrial environments, such as uneven illumination, noise and other factors, difficult to accurately identify the characters, performance of character recognition system of hardware and software also put forward higher requirements. This paper is based on the embedded hardware platform with intelligent camera, bearing the pressed characters to recognize the object in the previous experience and on the basis of the results, proposed own solution, realize the automatic recognition of the pressed characters. Bearing the related contents include the following aspects:
1) bearing image capture and preprocessing. Based on the analysis of imaging characteristics of bearing characters pressed on the design of the image acquisition system. The system is suitable for bearing the embedded image acquisition platform based on smart camera, and the low angle LCD ring light source as auxiliary light source, through collecting image bearing pictures of the system can be confirmed to obtain high quality bearings. In the image preprocessing stage, because the bearing is a circle structure and character region in the image of the location is not fixed, so the need for positioning and correction of bearing character area, so that the subsequent processing and recognition. Therefore, the Otsu method is used in bearing two value image, first of all to the localization of the bearing circle. For lack of randomized Hough transform circle detection algorithm, using random circle detection algorithm is improved for the bearing center localization, in the selection of candidate circle In the process of judgment by simply reduces the amount of computation and experiments prove that the location algorithm can quickly and accurately locate the bearing circle. Then, in order to locate the center as the origin of polar coordinates, bearing character area is determined by projection method. Finally, the bearing character of sector correction, if using the polar coordinate transform the image of burr more, this paper uses character correction method based on affine transformation. The experimental results show that the method can transform the image of burr less and speed can meet the requirements.
2) bearing character feature extraction and classifier. This paper describes the extraction methods of two kinds of features: directional line element feature and contour feature, and on the basis of theoretical analysis and experiments, for their shortcomings are proposed based on elastic mesh directional line element feature extraction and contour level feature of wavelet transform two kinds of improved methods. Based on the characteristics of elastic mesh directional line element features effectively compensate for the sensitivity of uniform mesh deformation character based on robust feature. Characteristics of wavelet transform contour level features full use of the wavelet multi-resolution analysis based on the characteristics of better noise resistance and lower dimension on the bearing. Character recognition classifier, this paper discusses the principle of BP neural network and support vector machine classifier, and we design the momentum and adaptive learning rate based on BP The network and support vector machine network based on LIBSVM. The feature extraction method and classifier scheme adopted in bearing character recognition system are determined through the final recognition experiment.
Through experiments on real images, verify the validity of the method, and determine the selection of feature extraction and classifier. Experimental results show that the bearing image acquisition scheme used in this paper can obtain high quality, pretreatment method is better, according to the results of selected contour level features improved as features, support vector machine is used as classifier. Character recognition accuracy is above 96%, the speed can meet the practical demands. This method can quickly and accurately on the bearing characters pressed automatic identification.

【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP368.1;TP391.43

【参考文献】