摘要
本文提出了一种基于无约束手写数字的轮廓层次特征的神经网络识别算法。采用递推最小二乘BP(RLS—BP)训练算法能够有效地提高网络收敛速度,无需选取学习步长和动量因子,减少了权值矩阵及参数初值的影响。通过在网络结构中引入子类别输出节点,进一步降低网络的训练难度,增强了分类器的分类能力。将其应用于无约束手写数字的信函分拣系统,单字识别率达到98%以上,得到了令人满意的结果。
A new neural network classifier is proposed for unconstrained handwritten numerals based on the layer outline feature in this paper. Recursive least, squares BP (RLS-BP) training algorithm is considerably faster than the BP algorithm, and has an added advantage of being less affected by poor initial weights and setup parameters. By introducing sub-classes output layer nodes in neural network, RLS-BP algorithm will converge easier. Applying the classifier to the zip code recognition system, it is shown that the average recognition rate of single digit is above 98%. The experimental results are very satisfactory.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
1999年第4期416-423,共8页
Pattern Recognition and Artificial Intelligence
关键词
字符识别
特征抽取
神经网络
最小二乘算法
Optical Character Recognition, Feature Extraction, Neural Network
