基于神经网络的车牌识别技术研究

发布时间：2019-06-11 00:57

【摘要】： 车牌自动识别系统是以汽车牌照为特定目标的专用计算机视觉系统。它的研究主要涉及到了模式识别和人工智能、计算机视觉、数字图像处理、人工神经网络等众多的学科领域。本文通过对车牌识别系统中车牌定位、图像预处理、倾斜矫正、字符分割、字符识别五个关键环节的分析研究,设计了一个完整的车牌识别系统,并在MATLAB环境下进行了仿真模拟。在车牌的定位部分,本文采用的是基于颜色特征和纹理特征的车牌定位方法。首先将彩色图片从RGB空间转换到HIS空间,利用蓝底白字车牌中蓝色的色度和饱和度S值较大的特点,实现了车牌的粗定位。然后再对粗定位后的图像利用Canny算子进行边缘检测,根据车牌部分图像黑白跳变频率较高的特征,最终实现了车牌的精确定位。在图像的预处理部分,本文将得到的车牌定位图像进行了灰度化处理,利用Otsu法将灰度图像转换为二值图像,并给出了一种灰度图像增强算法,对采集到的车牌图像进行增强处理;由于在实际中车牌的边框和上下铆钉会对车牌的识别工作形成干扰,因此在该部分中对车牌的边框和铆钉进行了去除。在获取车辆图像的过程中,由于摄像机和车牌之间角度的变化,经常使所拍摄的车辆图像发生倾斜,导致车牌扭曲和字符变形,给字符分割和字符识别带来极大影响。为此,文章研究了一种基于空间扭曲校正和Hough变换的车牌图像校正方法。在字符的分割部分,本文依据现行的车牌设计原则,利用改进后的水平投影法,将车牌图像分割7个待识别字符,并对分割后的字符进行了归一化处理。实践证明该方法对解决汉字的不连通问题、字符的粘连问题、噪声的干扰问题以及车牌的前2个字符和后面5个字符之间存在的小圆点问题是行之有效的。在字符的识别部分,采用改进后的BP神经网络,针对汉字、字母、字母或数字、数字四种不同的识别问题,设计了四种不同的分类器。利用13特征提取法进行特征提取,将其结果作为网络的输入,最后将不同的识别结果组合得到车牌号码。
[Abstract]:Automatic license plate recognition system is a special computer vision system with automobile license plate as a specific target. Its research mainly involves pattern recognition, artificial intelligence, computer vision, digital image processing, artificial neural network and many other disciplines. Based on the analysis and research of license plate location, image preprocessing, tilt correction, character segmentation and character recognition in license plate recognition system, a complete license plate recognition system is designed in this paper. The simulation is carried out in MATLAB environment. In the part of license plate location, this paper adopts the license plate location method based on color features and texture features. Firstly, the color picture is converted from RGB space to HIS space, and the rough location of license plate is realized by using the characteristics of blue chromaticity and saturation S in the license plate with blue background and white character. Then the edge detection of the rough positioning image is carried out by Canny operator. According to the high black and white jump frequency of some images of the license plate, the accurate location of the license plate is finally realized. In the part of image preprocessing, the license plate location image is grayed out in this paper, and the gray image is converted into binary image by Otsu method, and a gray image enhancement algorithm is proposed. The collected license plate image is enhanced. In practice, the border and rivets of license plate will interfere with the recognition of license plate, so the border and rivet of license plate are removed in this part. In the process of obtaining vehicle images, due to the change of the angle between the camera and the license plate, the vehicle image is often tilted, which leads to the distortion of the license plate and the deformation of the characters, which has a great impact on character segmentation and character recognition. In this paper, a license plate image correction method based on spatial distortion correction and Hough transform is studied. In the part of character segmentation, according to the current license plate design principles, this paper uses the improved horizontal projection method to segment seven characters to be recognized, and normalizes the segmented characters. It is proved by practice that this method is effective to solve the problem of disconnectedness of Chinese characters, adhesion of characters, interference of noise and the problem of small dots between the first two characters and the last five characters of license plate. In the recognition part of characters, four different classifiers are designed for four different recognition problems of Chinese characters, letters or numbers and numbers by using the improved BP neural network. The 13 feature extraction method is used for feature extraction, and the results are used as the input of the network. Finally, different recognition results are combined to get the license plate number.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2009
【分类号】：TP391.41

【引证文献】