【作者】 代青；

【导师】 戴亚文；

【作者基本信息】 武汉理工大学 ， 凝聚态物理， 2009， 硕士

【摘要】 水表智能识别系统是计算机视觉在日常生活中的一个重要应用,而图像处理和字符识别技术是其中两个关键技术。在此论文中,我们根据真彩水表图像自身的特点,构建一个水表字符自动识别系统。该系统包括图像预处理、特征提取和分类器三个部分。图像预处理部分主要基于图像的HSV色彩空间。对于采集到的彩色水表表盘图像,分析图像自身特点,分别提取其V、S和H分量有用信息,并结合相关算法,对其进行了有效的预处理。将图像格式从RBG空间装换到HSV空间后:对于V分量,我们利用其对图像进行灰度化,用Canny算子检测图像边缘,并对所得边缘图像使用Hough变换来检测水表字符边框以获得倾角,然后根据旋转算法进行倾斜校正,读数区域的水平定位也由Hough变换检测出的平行直线所确定;对于S和H分量信息,在图像二值化中得到充分利用:我们提出彩色水表图像基于S分量阈值法与H分量阈值法融合的二值化方法和基于色彩聚类与H分量阈值法融合的图像二值化方法,并通过实验证明了这两种方法在彩色水表图像二值化中的有效性;并探讨了两种方法的优缺点,最终选取了基于色彩聚类与H分量阈值法融合的图像二值化方法。字符分割步骤,考虑到字符粘连情况的存在,采用双次分割办法:首先进行基于直方图的阈值分割,然后进行基于字符轮廓的字符粘连分割。最后对所得图像进行归一化处理。实验证明,对于我们采集到的水表图像,此种基于HSV空间图像预处理方法比之RGB空间有显著的优势。在字符特征提取阶段,讨论了字符的轮廓特征和统计特征提取的优缺点,并最终采用统计网格特征:即将每幅数字图像平均分割为9个面,统计每个面中数字点所占个数,然后与面积取商。字符识别器采用现在通用的BP网络:根据本系统图像特点,构建一个三层BP网络来识别数字;中间层神经元的个数由matlab模拟BP网络误差逼进结果来确定。使用采集的水表图像对所设计识别系统进行试验,对所得结果进行分析,对系统存在问题进行改进,最终获得具有较好的鲁棒性识别系统。更多还原

【Abstract】 Water meter automatic recognition system is an important application of pattern recognition and computer vision in our daily life;and the image processing and digits recognition are its two core technologies.In this paper,we will introduce a way to build a recognition system based on the features of the true color images,including image pre-processing,feature extraction and classifying.The image pre-processing mainly based on the HSV color space.The useful information of the three parameters(H,S,V) and the significant algorithm form a effective image pre-processing.We should first convert the format of the color images from RGB to HSV.Then we can detect the parallel through the algorithm of Canny and Hough to obtain the obliquity of the image to correct it,and this is based on the parameter of V.Also,we use the parallel to locate the digits region.In binarization step,the H and S parameter are useful:we propose the S parameter and H parameter features fusion ruler and the C-means and H parameter features fusion ruler to distinguish the digits from background.The two means are effective and both have its merit and defect,and the C-means and H parameter threshold fusion ruler is the final choise.Two major units compose the segment of digits:the pixels and the figure of the digits.Finally,we resize all the pictures into the same size.We choose the statistical features as the extraction features of digits.Four lines(including two horizontal and two vertical lines) divide the digit into nine regions.We total the number of the black dots to divide the total dots in each region as the features of the digits.Finally,a three-layer BP neural network is applied in this system to identify the number 0 to 9;the number of the neurons in the middle-layer of BP network is defined by the experiment result on matlab.We test the system by hundreds of water meter images available,and improve the system according to the experimental results.The results show the final proposed scheme is feasible and has great robustness.更多还原