基于改进的PSO神经网络的手写体汉字识别  

Handwritten Chinese Characters Recognition Based on Improved PSO Neural Networks

作　　者：郭志涛[1] 袁金丽[1] 张秀军[1] 范书瑞[1] 

机构地区：[1]河北工业大学信息工程学院,天津300130

出　　处：《河北工业大学学报》2007年第4期65-69,共5页Journal of Hebei University of Technology

摘　　要：为了克服手写体汉字识别中传统神经网络训练算法存在网络易于过早收敛的缺陷,本文提出采用自适应惯性权值的粒子群优化算法训练神经网络,即利用粒子更新迭代训练神经网络最优的权值和阈值,其中对粒子更新的惯性权值进行了自适应性的改进.试验中提取了大量汉字样本的笔画数量信息和分布信息作为特征向量,利用改进的粒子群神经网络对汉字进行分类,并与BP神经网络的识别效果做了比较.结果表明:自适应惯性权值的粒子群优化算法能有效避免网络"早熟",大大提高了网络训练精度,网络对汉字的识别正确率明显提高.In order to eliminate the shortcomings of traditional neural networks in handwritten Chinese characters recognition, such as the premature convergence, a novel intelligent method is presented, which uses the particle swarm optimization （PSO） algorithm with adaptive inertia weight to train the neural networks. The main idea is that the optimum weights and thresholds of the neural networks is acquired by the iteration and updating of the swarms, in this process, the inertia weight of the swarm iteration is improved to be adaptive in this paper. In the experimentation, the quantity and distribution information of the strokes of the Chinese character is extracted as the features, then the Chinese characters is classified by the improved PSO neural networks based on these features. Comparing with the BP neural networks, the improved PSO neural networks can avoid the premature convergence and achieve higher precision, in handwritten Chinese characters recognition, the application effect is very notable.

关 键 词：粒子群算法 惯性权值 神经网络 手写体汉字识别 图像特征提取 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]