【作者】 杨扬；

【导师】 李祥；

【作者基本信息】 贵州大学 ， 计算机软件与理论， 2006， 硕士

【摘要】 经过几十年的研究，人工神经网络目前已经广泛地用来解决模式识别和人工智能领域的一些复杂问题，并取得了以往传统算法无法获得的成功。光学字符识别(OCR)算法的研究属于模式识别领域的一个重要分支。很自然地，我们在光学字符识别算法设计中引入了人工神经网络。从上世纪70年代以来，光学字符识别技术逐渐走向成熟。对于印刷体字符的识别，目前已经有了很高的识别率；对于小规模的手写体字符识别，也已经走向实用。 相对于大规模字符识别(如汉字识别)系统，小规模字符识别系统实现难度相对较小，但在日常生活中却应用广泛。如阿拉伯数字识别系统，仅对10个阿拉伯数字和少量符号进行识别，可以广泛应用于邮政编码识别、汽车牌照识别，流水线上产品编号识别等领域。本课题选题的初衷，就是利用人工神经网络实现一套通用的小规模光学字符识别系统对邮政编码进行识别。 当前，由于半导体加工工艺的进步，微处理器的体积越来越小，速度越来越快。由这些微处理器组成的嵌入式系统，结构简单，成本低廉，被广泛应用于家电、工业控制等各个方面。正是出于生产成本和产品尺寸的考虑，本文的硬件模块将使用基于ARM系列微处理器的嵌入式系统取代传统的PC机。 本文的研究工作可以划分为硬件和软件两部分；所做工作、技术难点与技术创新如下： 1．研究ARM体系的嵌入式系统的设计。 a) 以Atmel公司的AT91RM9200处理器为核心，设计一套嵌入式系统。 b) 研究了高频电子印刷板的设计原理，并掌握了制作工艺。 c) 研究Linux操作系统代码，在嵌入式系统上移植了Linux操作系统。 2．实现基于人工神经网络的光学字符识别算法 a) 研究图像的获取、和预处理算法。 b) 研究K-L变换在字符特征抽取中的应用。 c) 利用美国邮政服务数据库，训练了BP网络的分类器。使用测试集，验证了我们的分类器对手写体数字的识别率为92％。 d) 研究了其他一些分类器，并同我们的人工神经网络分类器进行了性能比较。 本文所实现的光学字符识别系统，成本较低，体积较小。结构上具有一定的通用性，能够使用在各种图像处理、图像识别或监控应用中。今后，我们将进一步改进视频采集部分，提高视频图像的分辨率，将本文成果作为机器视觉系统产品进行市场推广。做论文期间，作者已在《计算机科学》上发表论文一篇。更多还原

【Abstract】 After many years research, the artificial neural networks already widely used for to solve many complex problems in pattern recognition and artificial intelligence domain, and has got the success which the traditional algorithm was hard to obtain. The research of OCR (Optical character recognition) is an important branch on pattern recognition domain. Very natural, we introduce the artificial neural networks into OCR algorithm design. Since last century 70’s, the OCR technology has gradually moved to maturely. For now, the performance of the printed character recognition is high, and the small-scale handwritten character recognition works very well in practice.Compared with the large-scale character recognition system (for example, the Chinese character recognition), the small-scale character recognition system is more easily to be realized, and it is widely used in daily life. For example, the Arabic numeral recognition system, only carries on the recognition to Arabic numeral, is applied in so many domain, such as zipcode recognition, the license plate recognition, product serial number recognition on assembly line, etc. The topic of our study is using artificial neural networks to realize the small-scale character recognition system.For present, as a result of the progress of semiconductor technology, the MCU become cheaper and quicker. Using these new MCU, we can easily design the high performance embedded system in industrial domain or household appliances. From the cost and the size consideration, we use ARM based embedded system to implement our OCR system.Our work may divide into two topics: hardware and the software. Our study and innovation in this paper as follow:1. We studied the architecture of the ARM based embedded system.a) Using the Atmel’s AT91RM9200, to design our embedded system.b) Studying the craft and design of the PCB in high frequence signal system.c) Studying the Linux operation system, transplanting the Linux system into our system.2. We realized the artificial neural networks based optical character recognition algorithm.a) Studying the image extraction, image-preprocessing algorithm.b) Studying the K-L transformation applied in character features extraction.c) Using the US postal service database, training our BP network based classifier. (After training, we obtain a Arabic numeral classifier with 92% correct rate)d) Studying other some classifiers compared the performance of those classifiers with our BP network classifiers.更多还原