【作者】 李建美；

【导师】 路长厚；

【作者基本信息】 山东大学 ， 机械电子工程， 2008， 博士

【摘要】 工业中广泛应用的一类标识字符,是直接采用机加工的方法制作在物体表面上的,标识字符与母体同色,且为立体字符。正是这个特点使得对它们的研究难以直接借用现有的字符识别研究成果,因而成为光学字符识别中的一个新的研究领域。随着制造业信息化水平的提高,对这类字符的自动识别也成为产品信息化管理的一个必然要求。本文在教育部资助下,开展了如下研究工作:设计了获取标牌压印字符图像的机器视觉系统。采集图像质量的高低是影响压印字符识别性能的关键,基于此,首先介绍了机器视觉系统的组成,分析了各个硬件的性能参数和选择方法,并根据标牌图像采集系统的性能特点选择了相应的硬件。分析了光照对标牌字符图像的影响,在一个机器视觉应用系统中,好的光源与照明方案往往是整个系统成败的关键。字符图像能否很好地反映字符的特征信息,照明方案起着关键的作用。研究了不同的照明方案对标牌压印字符成像的影响,同时研究了字符载体的反射特性对图像质量的影响。基于LED光源等机器视觉硬件建立了标牌图像采集系统,通过对大量不同材质和版面的标牌进行图像采集实验,最终确定了低角度LED光源构成的照明方案作为标牌图像采集的最佳照明方案。研究了标牌图像的光照处理方法。选择辐射度光照模型作为标牌图像的模型。针对标牌图像的特性,将多种光照处理方法应用于标牌图像进行处理。在此基础上,根据标牌压印字符系统的特点,提出了一种对光照处理效果对比方案。通过理论分析和实验验证证明,同态滤波算法可以作为一种有效的标牌图像光照处理算法。提出了利用图像分层技术对压印字符图像进行自动分割的方法。系统地研究了标牌图像压印字符的自动分割问题。根据标牌图像的特点,提出利用图像分层技术对标牌压印字符进行分割的方法。首先根据标牌图像灰度分布的特点将图像从最暗层到最亮层进行分层,对最暗层图像进行自适应二值化可得到标牌图像中的印刷文本图像。继而,基于连通域分析对印刷文本图像进行分析确定其所在区域,最后将确定的印刷文本区域从原始图像中去除,可得到目标区域,即仅含有压印字符区域的目标图像。经过对大量标牌图像进行分割实验证明,图像分层技术是一种适用于标牌压印字符分割的有效算法。研究了压印字符图像质量的评价及恢复算法。图像质量是影响字符识别系统精度的关键,对标牌图像中进行质量自动检测能有效地降低识别系统的拒识率和误识率。提出了利用五个质量参数来评价标牌图像质量,即黑色斑点因子、白色斑点因子、粘连字符因子、断裂字符因子和字符的高度、宽度和位置等参数,对标牌图像的质量进行了整体评价,定量地给出图像的质量等级。针对某些具有特定质量的标牌图像建立了相应的图像质量改善方法。实现了从单幅图像中自动提取出不同图像质量等级的字符,即实现了不同质量等级字符图像的分类功能,为下一步针对不同质量等级字符建立相应的图像处理算法奠定了良好的基础。利用Gabor滤波器提取压印字符的方向特征进行识别。提出了一种基于灰度图像和Gabor变换的压印字符特征提取新方法。设计了一组Gabor滤波器.然后对压印字符的灰度图像直接进行特征抽取,分别提取压印字符水平、竖直、左、右对角线等四个方向上的局部笔画特征,在此基础上,建立一种不变Gabor特征空间。经大量实验证明,这种不变Gabor特征具有优良的可分性,并且具有良好的旋转、尺度和平移不变性,同时对光照变化和噪声也具有较强的鲁棒性,从而进一步证明了Gabor滤波器适用于类似压印字符等低质量图像的特征提取。更多还原

【Abstract】 Characters formed on the surface of objects by machining method are widely used in industrial products. These characters are three-dimensional characters and they have the same color as the background. It is this characteristic that makes it difficult to use the existing research achievements of the traditional optical character recognition(OCR) to recognize them. Accordingly, recognition of such characters becomes a new hot spot in optical character recognition research field. Recently, with the development of the information technology, auto-recognition systems for pressed characters become a necessary requirement for Product Information Management. Supported by Doctoral Fund of Ministry of Education of China, the research carried out in this thesis is as followings.A machine vision image acquisition system for characters pressed on labels is designed. The quality of images acquired is a main factor to affect the performance of the pressed character recognition system. In order to improve the quality of the label image, the components of the machine vision system are introduced and their performance is analyzed. Illumination is vital in machine vision applications for providing contrast so that the camera can see the image clearly. Effect of the illumination on the label images is analyzed. To be effective, the illumination needs to provide an adequate amount of uniform light and be from the correct angle, so as to get enough information of the pressed characters from the label image. Effect of the several illumination schemes for machine vision imaging systems of label is studied. Also, the reflective properties of the label are analyzed. Based on LED lighting, an image acquisition system for label is established. Experiments on labels with different kinds of materials and layouts are carried out and the results prove that the low-angle LED lighting is the best illumination scheme for the characters pressed on label.Illumination preprocessing methods for label images are studied. According to the characteristics of the label images, several illumination preprocessing methods are applied on them. In order to choose a best illumination preprocessing method for label, a comparison scheme is designed. Both the theoretical analysis and the experiment results prove that the homomorphic filter can be reliably used for illumination preprocessing of label images.A novel segmentation method for characters pressed on label based on layered images is presented. Automatic segmentation methods for segmenting the characters from the label image are studied systematically. According to the characteristics of the label image, a segmentation method based on layered images is presented. The algorithm classified the label image into three planes: the brightest plane, the darkest plane and the middle plane. The printed text on the label is extracted from the adaptively binary image of the darkest plane. Then, the block of printed text is defined based on connect components analysis. Finally, by removing the block of the printed text from the original image, the pressed characters are picked up successfully. Experiments are carried out with a large number of label images. Experiment results show that the proposed segmentation method can be reliably used in the label images.Image quality assessment and restoration algorithms for the pressed characters are studied. The quality of image acquired is a key factor to affect the accuracy of character recognition system. Algorithms for estimating the quality of any given images can be used to automatically determine improvement made by adaptive image-restoration algorithms. In order to estimate the quality of the image, five quality measures are introduced. That is, small speckle factor, white speckle factor, broken character factor, touching character factor and character-height, width and location factor. Based on the information provided by the quality measures, a quantitative grade can be given to the label image for its quality estimation. Several image-restoration algorithms are introduced to improve the label image. Also, based on the quality measures, characters with different qualities can be segmented from a label image, which is helpful to adaptly select the image processing algorithm for characters with specific quality. Recognition the pressed characters with their direction feature extracted by Gabor filters. A novel feature extraction method for raised or indented characters pressed on a label using Gabor transform based on the gray image is presented. A set of Gabor filters is designed according to the characteristics of the pressed characters. These filters are used to break down a gray image of a character into four directional feature sub-images to get the local Gabor features. Based on these local features, an invariant feature space is established. Experiments are carried out with a large number of raised or indented character images. Experiment results show that the proposed invariant Gabor features have great separate capability and have great robust to rotation, scale, translation invariance and are also insensitive to illumination conditions and noise changes. It is proved that Gabor transform can be reliably used for extracting the feature of low quality character in image processing.This work is supported by Doctoral Fund of Ministry of Education of China.更多还原