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。 This paper presents a neural network recognition algorithm based on outline features of unconstrained handwritten numbers. Recursive Least Squares BP (RLS-BP) training algorithm can effectively improve the speed of network convergence without the need to select the learning step and momentum factor, and reduce the influence of weight matrix and initial parameters. By introducing sub-category output nodes in the network structure, the training difficulty of the network is further reduced, and the classification ability of the classifier is enhanced. Applying it to the letter sorting system of unconstrained handwritten numbers, the recognition rate of single word reaches more than 98%, and satisfactory results have been obtained.