【作者】 万忠；

【导师】 张铃；

【作者基本信息】 安徽大学 ， 计算机应用技术， 2006， 硕士

【摘要】 模式识别是计算机智能的一个重要研究领域，它在医学、工程、交通、天文、军事等很多领域有着广泛的应用。 本文以轨道衡车牌识别系统为应用背景，对车牌识别系统的各阶段的任务及遇到的问题进行了比较详细的说明和分析研究，然后基于构造性覆盖算法和商空间粒度计算理论提出了一种多粒度的字符识别覆盖算法，并且基本完成了轨道衡车牌识别系统的设计与实现。本论文所做的主要工作如下： 1．针对所处理车牌的特点，基于对大量车牌灰度图的直方图分析，我们采用灰度拉伸等操作去除了背景中干扰目标的象素信息，成功地进行了预处理，提高了车牌定位的准确率。 2．在车牌的定位阶段，我们通过对水平纹理投影平滑后的峰谷分析，对粗定位的上下边界进行了微调；在字符分割阶段中，我们通过对目标象素的竖直投影平滑后的峰谷分析，并且利用车牌字符的宽、高等先验信息改善了字符分割的效果；在车牌去噪和倾斜校正阶段，我们利用连通域标记算法都取得了较好的效果。 3．从分类器的结构方面考虑，我们提出了一种基于覆盖算法的两层结构分类器的设计方法，即多粒度的字符识别覆盖算法，然后将它和单层结构分类器做了实验分析对比，得出了在不明显增加构造复杂度的情况下两层结构的设计大大改善了分类器的性能。 本文的轨道衡车牌识别系统是模式识别领域的一个典型应用。它的基本思路和具体设计可扩展适用于其他行业类似的分类应用中，具有很好的应用前景。更多还原

【Abstract】 Pattern recognition is an important research field in computer intelligence, with its extensive application in medicine, engineering, traffic, astronomy, military area and so on.This thesis takes as the application background the train-plate recognition system on the railway weighing-machine, comparatively explains the tasks and problems and does analytic research across all the phases of the system, and then puts forth an new algorithm namely multi-granular covering algorithm for character recognition based on the constructive covering algorithm and the quotient space granular computation theory, furthermore generally completes the design and implementation of the system. The primary work of the thesis is listed below:1. Considering the characteristics of train plate treated with, we adopt gray-extending and other operations to eliminate the pixels that disturb the targeted area based on the histogram analysis of a large number of train-plate gray bitmaps, and carry out pre-processing successfully for original train-plate bitmaps, and thus increase the correct ratio of plate locating operation;2. In the locating phase of the system, through analyzing the valleys of the smoothed horizontal gray texture projection, we perform minor-adjusting to the bottom and top position determined by the rough locating operation; in the phase of character segmentation, we improve the operation by analyzing the valleys of the smoothed projection of the targeted pixels from left to right and by using the width, height and other basic imformation of the processed plate character; in the denoising and inclination-adjusting phase of the system, we obtain good results by using the connected region labelling method;3. Taking the structure of classifier itself into consideration, we puts更多还原