基于改进BP神经网络的手写字符识别

Handwritten Character Recognition based on Improved BP Neural Network

针对标准反向传播(BP,Back Propagation)神经网络算法收敛速度慢、易陷入局部极小等缺点,采用附加动量法与学习速率自适应调整相结合策略对神经网络初始参数进行设置。通过在权重计算公式中加入动量项,降低神经网络对误差曲面局部调节的敏感性,有效抑制其陷于局部极小。学习速率根据总误差的变化进行自适应调整,可以有效地缩短学习时间,加快收敛速度。将该改进算法应用于数字、英文字母以及简单汉字的手写字符识别系统中,进行了有无动量、有无噪声等实验,结果表明该方法与传统BP算法相比识别精度较高、训练时间较短且具有较强的鲁棒性。

A modified back propagation（BP） algorithm with additional momentum and self-adaptive learning rate is adopted to set up the initial parameters of the neural network,which could overcome the disadvantages of standard BP algorithm,such as slow convergence rate and easy fall into local minimum.The network sensibility to the local adjustment of error curved surface could be reduced and the local minimum suppressed efficiently by adding momentum item into the weight calculation formula,a self-adaptive learning algorithm is implemented,thus to adjust the learning rate according to the total error and shorten the training time and speed up the convergence rate.The improved algorithm is applied to recognising the handwritten characters,including numbers,English letters and some simple Chinese characters.The experimental results indicate that the proposed algorithm could achieve higher recognition accuracy,shorten the training time and has better robustness,as compared with the traditional BP algorithm.