【作者】 蒋伟；

【导师】 古钟璧；

【作者基本信息】 四川大学 ， 模式识别与智能系统， 2003， 硕士

【摘要】 手写数字识别(Handwritten Digits Recognition，简称HDR)是光学字符识别技术(Optical Character Recognition，简称OCR)的一个分支，它在邮政编码自动识别、银行、金融及模式识别理论研究等众多领域有着广泛的应用前景。 人工神经网络(Artificial Neural Network，简称ANN)的兴起源自于人类对自身的模仿。1943年出现了神经元的数学模型，随后出现的学习算法进一步推动了对神经网络的研究，而且逐渐地从功能模拟为目标转向基于知识处理的应用研究。BP神经网络由于其算法简单而且有效而在实际中应用广泛。 遗传算法(Genetic Algorithm，简称GA)是一种基于达尔文自然进化论和孟德尔遗传变异理论的高效并行全局搜索方法，它能在搜索过程中自动获取和积累有关搜索空间的知识，并自适应地控制搜索过程以求得最优解。 本文从神经网络和遗传算法基本理论及算法构成入手，将神经网络用于手写数字识别中。针对BP神经网络易陷于局部最优解、依赖初始权值等缺点，提出了用遗传算法取代BP神经网络中的BP学习算法。并结合实际的手写数据库，对该算法和BP算法进行了比较。 试验结果表明：基于遗传算法的神经网络用于手写数字的识别是可行的和有效的，该算法能够避免陷于局部最优解，不依赖于网络初始权值，也能达到较高的识别率。更多还原

【Abstract】 Handwritten digits recognition is one branch of optical character recognition (OCR), which has broad potential application in fields of postal service, banking, finance and pattern recognition theory research.Artificial neural network originates from human’s imitation to self. The mathematical model of neural cell was presented in 1943 and the following advance of learning algorithm encouraged the research of neural network. Backpropagation (BP) learning algorithm is commonly used in practical use because of its simple and efficiency.Genetic Algorithm (GA) is a kind of parallel and efficient search algorithm for global optimal solution, which is based on Darwin’s natural evolution theory and Mendel’s genetic mutation theory. GA has the ability of automatic acquiring and storing knowledge of search space in the process of searching optimal solution and can adaptively control the search process to approach the optimal solution.In this thesis, the author based on the basic theory of neural network and genetic algorithm and applied genetic algorithm based neural network to handwritten digits recognition. Seeing that BP neural network may easily get struck into local optimal solution and is dependent on the initial weights, the author proposed to replace the BP learning algorithm with genetic algorithm to train the neural network and then compared this two algorithms in the application of handwritten digits recognition.Experimental result shows that the genetic algorithm based neural network is efficient for handwritten digits recognition, because this method can avoid to be struck into local optimal solution, does not dependent on initial weights and can also achieve high recognition rate.更多还原