基于自组织特征映射神经网络的数字模式识别

在分析自组织特征映射(SOFM)神经网络基本学习算法的基础上．从提高算法收敛速度和性能出发．提出了一种改进算法：随机选择样本输入次序；根据实际应用并结合专家经验确定初始连接权值；采用高斯函数作为拓扑邻域函数；将算法分成排序和收敛两个阶段。并分别采用不同的学习率和邻域函数．采用改进后的SOFM算法对输入样本进行自组织聚类，再利用学习矢量量化(LVQ)算法解决样本分类中的交迭问题。提高了分类精度．仿真实验结果表明．该网络能够识别常用的数字(0～9)和英字母．特别是在有噪声污染的情况下．可以获得较好的效果。

Digital Pattern Recognition Based on Self-organizing Feature Map Neural Network

An improved algorithm for self-organizing feature map neural network was presented in the paper.In this algorithm,the samples were ordered at random for improving the learning efficiency and the expert experience and actual state were considered using training weight vectors.The Gauss neighborhood function was used to replace the square or circular function,and different descending functions of learning rate and neighborhood width were used in two learning periods.The hybrid of the SOFM algorithm and LVQ algorithm can implement the supervised learning of the SOFM network.Simulation experiments indicate that the network can recognize numeral symbols (0~9) and English characters,and gain better benefit when the symbols and characters are polluted.