最新基于Hopfield神经网络的字符识别.doc

基于Hopfield神经网络的字符识别 摘要: 文章介绍了离散Hopfield神经网络的基本概念及其原理，以Matlab为工具，根据Hopfield神经网络的相关知识，设计了一个具有联想记忆功能的离散型Hopfield神经网络。首先提取照片的像素值，通过对照片的灰度处理，得到灰度像素矩阵。由于对单个字符进行识别的效果比多个字符整体识别的效果好，故对不同的字符进行分割，然后运用OSTU算法求得最佳阈值，通过数据替换得到该字符的二值矩阵。用原图片的二值矩阵作为训练样本，生成Hopfield神经网络。然后分别在不同噪声强度的情况下，以噪声图像的二值矩阵作为测试样本，观察网络的输出效果，并计算出相应的识别率。通过测试发现，噪声强度在较小范围0.1左右时，该网络可达到很好的识别效果，此时识别率接近1；随着噪声强度的增大，识别效果变差；当噪声强度达到0.4时，该网络已无法进行识别。 关键字： Hopfield神经网络 二值矩阵 OSTU算法 识别率 Character recognition based on Hopfield neural network Abstract: the article introduces the basic concept and principle of discrete Hopfield neural network,which is based on the Matlab tools, and Hopfield neural network knowledge, it designed a discrete Hopfield neural network with associative memory function. First, extract the image pixel values, through processing the image grayscale , gray pixel matrix is obtained. Due to the effect of single character recognition is better than characters overall recognition , so the different characters need segmentation. Then optimal threshold is obtained by the use of OSTU algorithm, and the binary matrix of the characters is created by replacing data. Two-valued matrix in the original image is regarded as the training sample, and it generates the Hopfield neural network. Then respectively in the case of different noise intensity, treating binary matrix of the noise image as test samples, compare the effect about output of the network , and calculate the corresponding recognition rate. Through test, found that when the noise intensity in small within 0.1, the network can achieve good recognition rate, and recognition rate at this time is close to 1; with the increase of noise intensity, recognition rate is lower; while the noise intensity is 0.4, the network has been unable to identify. Key words: Hopfield neural network two-valued matrix OSTU algorithm recognition rate 一、原理概述 1Hopfield网络的拓扑结构 Hopfield最早提出的网络是二值神经网络，神经元的输出只取1和-1，所以也称离散神经网络（DHN