【作者】 赵晓霞；

【导师】 方敏；

【作者基本信息】 合肥工业大学 ， 控制理论与控制工程， 2005， 硕士

【摘要】 由波兰学者Z.Pawlak提出的粗糙集理论,是一种研究不精确、不确定性知识的数学工具。利用粗糙集理论,可以方便地描述知识表示中不同属性的重要性,减少知识表示空间的维数,但其模式识别基于规则匹配方法,对噪声敏感,泛化能力不强。而神经网络以其优越的泛化能力和容错能力,广泛用于模式识别中。二者的结合既可以简化信息处理的复杂性,又可以提高信息处理的精度。汽车牌照的自动识别是机器视觉与模式识别技术在智能交通领域的重要应用,进行基于粗糙集与神经网络结合的车牌识别系统的研究具有一定的理论意义和实践价值。本文结合粗糙集与神经网络理论,用于车牌识别系统,完成了以下工作:1. 实现了四种图像二值化方法:类内最小交叉熵法、类间最大交叉熵法、类内最小模糊散度法和类间最大模糊散度法。2. 在车牌定位方法的研究上,提出了一种基于颜色分析的车牌定位方法,对车牌的大小、汽车在图像中的位置以及图像背景的限制较少,定位效果好。3. 在基于粗糙集的知识获取的研究上,提出了两种属性离散化算法:基于NCL聚类学习方法的属性离散化算法和基于自组织特征映射网的属性离散化算法;提出了三种改进的属性约简算法:改进的代数集合算法、重要度加权平均算法和改进的可辨识矩阵算法。4. 在车牌字符识别的研究上,利用粗糙集理论的知识获取和知识约简能力,以及神经网络的学习能力和泛化能力,构造车牌字符识别系统。该系统既简化了识别系统的结构,又提高了知识的泛化能力。在此基础上,将粗糙集、神经网络和模糊逻辑相结合,提出了一种粗糙模糊神经网络识别器,可以根据粗糙集获取的知识进行模糊化后构造神经网络,有效地提高字符识别的抗干扰性和识别率。更多还原

【Abstract】 Rough Set theory, a new mathematical tool dealing with vagueness and uncertainty, was introduced by Poland scholar Z. Pawlak. It can describe the significance of different attributes and reduce the dimension of knowledge space. But its pattern recognition based on rule matching is sensitive to disturb. Neural network has better robustness and widely used in pattern recognition. Combining these two theories not only simplifies the complexity of information processing but also improves the precision of information processing. Vehicles’ License Plate Recognition (LPR) system is an important application of computer vision and pattern recognition in the intelligent transportation management. So, the research on LPR system based on rough set and neural network is of certain academic and practice value.In this article, rough set and neural network are combined and used in the LPR system. Some work has finished as following.Firstly, four binarization algorithms are presented, including minimum cross-entropy, maximum between-class cross-entropy, minimum fuzzy divergence and maximum between-class fuzzy divergence. Secondly, an effective locating algorithm is presented based on color analysis, which has fewer limits to the car size, the car position in the image and the image background. Thirdly, two attribute discretization algorithms are presented, which are based on NCL clustering method and self-organizing feature map (SOFM) neural network respectively. And three modified attribute reduction algorithms are presented, including modified algebraic algorithm, weighed sum of attribute significance algorithm and modified discernible matrix algorithm. Fourthly, two effective character recognition methods are presented. In the first method, rough set is used to solve the knowledge acquisition and reduction, and neural network is used to solve the pattern recognition. In the second method, a rough fuzzy neural network classifier is introduced, which combines the rough set, neural network and fuzzy logic. It can efficiently improve the robustness and recognition rate of the LPR system.更多还原