基于遗传优化DHNN的噪声图像识别研究  

Study on Noisy Image Recognition using Optimized DHNN by Genetic Algorithm

作　　者：郭勇[1] 乐江源[1] 李梦超 郑景林 GUO Yong;LE Jiangyuan;LI Mengchao;ZHENG Jinglin(School of Physics and Electronic Information,Gannan Normal University,Ganzhou 341000,China)

机构地区：[1]赣南师范大学物理与电子信息学院,江西赣州341000

出　　处：《赣南师范大学学报》2022年第6期43-47,共5页Journal of Gannan Normal University

基　　金：江西省教育厅科技项目(GJJ201437)。

摘　　要：采用外积法设计的离散型Hopfield神经网络(DHNN),系统稳定但识别噪声图像能力有限,因此引入遗传算法对网络参数进行优化.首先将权值与阈值编码形成染色体,创建初始种群;然后对种群中个体的染色体进行选择、交叉、变异操作,将适应度值小的个体替换同代种群中适应度值大的个体,形成子代种群,并按照同样的方式遗传若干代;最后将末代种群中适应度值最小的个体的染色体解码为权值和阈值.对3种交通标志噪声图像识别,优化前,当噪声水平增大至0.3时识别效果较差,且PSNR随着噪声水平的增大而减小;优化后识别效果更好,且PSNR先增后减,在0.5时最低.因此,采用遗传算法优化后DHNN识别噪声图像能力更强,然而算法收敛时间长,需要进一步完善.When the outer product method is used to design the network parameters of Discrete Hopfield Neural Network(DHNN),the system is stable,but noisy image recognition ability is limited.Therefore,genetic algorithm is introduced to optimize the network parameters.Firstly,the weights and thresholds are encoded to form individual chromosomes,so as to create the initial population;secondly,the chromosomes of individuals in the population are selected,crossed and mutated,and the individuals with large fitness values are replaced by the ones with small fitness values in the same generation population to form the offspring population,and the offspring population is inherited in the same way for several generations;lastly,the individual with the smallest fitness value in the last generation population is decoded into weights and thresholds.Three noisy traffic sign images are cited as instances,before optimization,the recognition effect is poor when the noise level increases to 0.3 and PSNR is decreased with the increase of the input noise level,while after optimization,the recognition effect is better,PSNR is increased first and then decreased,with the lowest value under noise level 0.5.In conclusion,after using genetic algorithm,the network's ability to recognize noisy images is enhanced.However,the genetic algorithm needs further optimization because of its longer convergence time.

关 键 词：噪声图像识别 DHNN 权值 阈值 遗传算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]