车牌图像预处理与字符分割算法研究

发布时间：2018-11-24 14:29

【摘要】：自动车牌识别是智能交通系统中的重要模块,一般包括车牌区域提取、车牌字符分割、车牌字符识别等处理步骤。由于受光照不均、字符阴影、背景纹理等干扰的影响,字符分割和识别的精度较低。因此,本文提出车牌字符分割及其预处理技术研究的课题,开展该项研究对提高车牌自动识别系统的精确性和鲁棒性有重要的应用价值。本文主要工作和贡献如下:基于车牌区域垂直边缘高密度、边缘点邻域灰度对相似等特点,提出了一种抗背景区域干扰的车牌水平校正算法。首先,通过垂直边缘密度滤波、近邻垂直边缘水平连接,得到车牌区域块,并根据字符高度一致性以及位置的渐变性,提取有效的字符列,然后,利用最小二乘法基于所有字符列中点拟合直线,估计出车牌区域倾斜角。基于车牌底颜色模式种类固定、字符笔画宽度一致的性质,提出了抗光照变化和背景纹理干扰的车牌颜色模式判断算法。通过分析车牌区域可能出现的颜色并利用字符宽度一致性的特点,对色彩饱和度高和饱和度低的车牌分别使用颜色量化直方图和形态学处理的方法来判断。基于车牌区域字符面积比例固定和颜色一致等特点,构建了一种自适应的车牌二值化方法。首先,计算Otsu全局阈值,基于光照的平滑变化,通过与Bernsen局部阈值的加权,得到灰度接近全局阈值的像素的阈值,来克服光照不均的影响,再通过前景所占比例判断是否需要选用基于RGB颜色空间的k-means聚类方法来区分字符阴影或非车牌区域、车牌背景和字符区域。基于车牌底色一致和字符排列方式固定等特点,提出了一种抗背景区域干扰的水平车牌字符分割方法。首先,对车牌二值图垂直投影进行波谷分析,利用车牌底色的一致性估计左右边界,并找到第二、第三个字符之间的间隔位置,为模板匹配提供参考位置;设计变长模板,在垂直投影直方图上滑动,结合波谷约束以最大类间方差准则找到最佳匹配参数,最后依据投影和字符宽度一致性进行分割位置的精确调整。最后,设计了本文算法的仿真实验。对于现有700幅车牌图像的数据集(包含无干扰和有光照影响、图像背景区域干扰的图像),分割准确率达到了97.28%,表明本文提出算法具有较好的适应性。
[Abstract]:Automatic license plate recognition is an important module in intelligent transportation system, which generally includes the processing steps of license plate region extraction, license plate character segmentation, license plate character recognition and so on. The accuracy of character segmentation and recognition is low due to the influence of uneven illumination, shadow of characters, background texture and so on. Therefore, this paper puts forward the research topic of license plate character segmentation and its preprocessing technology, which has important application value to improve the accuracy and robustness of license plate automatic recognition system. The main work and contributions of this paper are as follows: based on the high density of vertical edge of license plate area and the similarity of adjacent gray pairs of edge points, a license plate level correction algorithm against background interference is proposed. First, through vertical edge density filtering, the nearest vertical edge is connected horizontally, and the license plate area block is obtained. According to the character height consistency and the gradual change of position, the effective character column is extracted. The least-square method is used to estimate the inclination angle of license plate area by fitting straight lines based on the midpoints of all character columns. Based on the fixed color pattern at the bottom of the license plate and the uniform width of the character stroke, an algorithm for judging the color pattern of the license plate is proposed, which can resist the interference of illumination and background texture. By analyzing the possible colors in the license plate area and using the characteristics of character width consistency, the color quantization histogram and morphological processing are used to judge the license plate with high color saturation and low saturation respectively. Based on the characteristics of fixed area and uniform color in the license plate area, an adaptive binarization method of license plate is proposed. First of all, the global threshold of Otsu is calculated. Based on the smooth change of illumination, the threshold of pixels with gray level approaching global threshold is obtained by weighting with local threshold of Bernsen to overcome the influence of uneven illumination. Then the proportion of foreground is used to determine whether it is necessary to choose k-means clustering method based on RGB color space to distinguish character shadow or non-license plate region, license plate background and character region. Based on the characteristics of uniform background color and fixed character arrangement, a method of horizontal license plate character segmentation against background interference is proposed. Firstly, the vertical projection of the binary image of the license plate is analyzed, and the left and right boundary is estimated by the consistency of the bottom color of the license plate, and the interval between the second and the third characters is found to provide the reference position for template matching. The variable length template is designed to slide on the vertical projection histogram and the maximum inter-class variance criterion is used to find the best matching parameter combined with the trough constraint. Finally the segmentation position is accurately adjusted according to the consistency of projection and character width. Finally, the simulation experiment of this algorithm is designed. For the existing data sets of 700 license plate images (including images with no interference, illumination, background area interference), the segmentation accuracy reaches 97.28, which shows that the proposed algorithm has good adaptability.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】