离散Hopfield神经网络的手写数字识别研究  

Study of Handwritten Digital Recognition in Discrete Hopfield Neural Networks

作　　者：潘园园[1] 张力 段玲玲 段法兵[1] PAN Yuanyuan;ZHANG Li;DUAN Lingling;DUAN Fabing(Institute of Complexity Science,Qingdao University,Qingdao 266071,China)

机构地区：[1]青岛大学复杂性科学研究所,山东青岛266071

出　　处：《复杂系统与复杂性科学》2018年第1期75-79,共5页Complex Systems and Complexity Science

基　　金：国家自然科学基金(61573202)

摘　　要：利用离散Hopfield神经网络对手写数字识别进行了研究。将受到噪声污染的手写数字储存为二值图像,然后调制成二进制信号通过神经网络进行传输,通过给定权矩阵的Hopfield神经网络进行按址存储,将网络输出的内容再映射为数字图像。实验结果表明:数字图像识别的误码率与调制信号的幅值、码间时间间隔和网络神经元耦合个数成负相关关系,而且随着噪声强度的增加,误码率出现非周期随机共振现象,在一非零最优噪声强度值达到最小,此时数字图像也恢复得更加清晰。这些结果为进一步研究最小误码率优化目标下的Hopfield神经网络自适应权重矩阵提供了实验依据,而且对于神经网络联想记忆中随机因素的作用研究具有重要意义。This paper studies the handwritten digital recognition by the discrete Hopfield neural networks.In the experiment,the noisy handwritten digital image is transferred into the binary signal by the serial-scan mode.The binary modulated signal is transmitted through the neural network with the designed weight matrix and the output storage mode of the network is mapped into digital image.The error rate of the digital image is negatively correlated with the amplitude of the modulated signal,the time interval and the number of coupled neurons in the network.However,as the noise intensity increases,the error rate manifests the aperiodic stochastic resonance effect,and achieves the minimum at the non-zero optimal noise intensity.Under this circumstance,the recovered digital image appears more clearly.These results provide a theoretical basis for further research on the adaptive weight matrix of Hopfield neural network for obtaining the minimum error rate,and also are of significance for the positive role of the randomness in the associative memory of the neural networks.

关 键 词：离散HOPFIELD神经网络 数字图像识别 误码率 随机共振 

分 类 号：TN911.7[电子电信—通信与信息系统] N945.12[电子电信—信息与通信工程]