【作者】 王志红；

【导师】 王建平；

【作者基本信息】 合肥工业大学 ， 检测技术与自动化装置， 2003， 硕士

【摘要】 小波变换具有“显微镜”特性和类人视觉特点，在图像处理及模式识别中有着越来越广泛的应用；神经网络模式识别表现出来较强的自学习、自适应能力以及容错性、鲁棒性等使得它成为模式识别领域一个重要的研究方向。 车牌识别是计算机视觉与模式识别技术在智能交通领域的重要应用，是实现交通管理智能化的核心技术。因此，将小波变换与神经网络有机结合并运用于车牌自动识别具有较大的理论意义和实践价值。 论文主要涉及了以下工作： 1)图像预处理。该阶段研究图像的消噪和倾斜度校正。通过分析图像的小波分解高频系数特性，提出了一种小波局部阈值消噪方法。并且应用简化改进的Hough变换，通过检测边界线段检测牌照倾斜角度，对车牌进行倾斜校正。 2)字符分割。提出了一种新的实用车牌字符分割方法。根据车牌的先验知识，利用小波变换自动变焦、多尺度分析的特点，对传统的垂直投影法及多线垂直投影法进行滤波，自适应搜索分割点，高效地进行了车牌字符分割。 3)特征提取。利用图像小波变换的方向性分解构造出针对易混淆字符的细小波网格特征提取方法。 4)字符识别。对多特征、多神经元网络集成识别方法进行了研究，从而提出了一种基于神经网络的多级串行分类器的车牌字符识别方法。 本文研究表明，小波局部阈值降噪法有效可行；字符分割算法对前续车牌定位及倾斜程度要求低，抗干扰性及适应性强；小波网格特征稳定性好，区分性强，多级网络分类器针对性及泛化能力强，与单一网络分类器相比，可有效提高系统的抗干扰性和识别率。更多还原

【Abstract】 Wavelet transform has been widely used in image process and pattern recognition because its "micro - scope" feature and similar-human vision feature. Neural net pattern recognition is an important research direction in pattern recognition field because of the strong self-learning, application feature.VLP is an important application and the core technology of computer vision and pattern recognition in the intelligent traffic field. So apply these two tools in VLPR has biggish theory meaning and practice value.This paper has finished some tasks as follows:1) image pretreatment. This phase research image de_noise and image inclination rectification. Based on the wavelet high frequency coefficient, we bring forward a wavelet part thresholding method. While in image inclination rectification, a modified Hough transform is used in VLP inclination detection and rectification.2) character segmentation. This phase bring forward a new applied method in VLP character segmentation. Based on predeterminate knowledge of VLP and wavelet nulti-scale analysis feature, segmentation position is located self-adaptively.3) character feature extraction. According to wavlet’s directive characteristic,we bring forward a new wavelet part grid feature.4) character recognition. Study the integration method of many inputs and many neural net. Therefore we bring forward a multi-layer serial classifiers with neural network.The research in this article shows that the wavelet part thresholding method is effective; character segmentation method has little limitation for the VLP inclination; the wavelet grid feature has good stability and satisfactory distinction; compared with the single classifier, the multi-layer serial classifiers can efficiently improve resistance to interaction and recognition rate of the system.更多还原