基于PSO改进的BP神经网络字符识别算法研究

Research on Improved BP Neural Network Character Recognition Algorithm Based on PSO

摘要: 针对字符识别算法中常规的BP神经网络算法在训练识别多种类字符时易造成过拟合而导致训练失败，字符识别率不高，传统的模板匹配法对字符识别的效果也不好的问题，提出一种基于PSO改进的BP神经网络字符识别算法。采用粒子群优化算法(PSO)计算BP神经网络的初始权值和偏置，提高了网络的收敛速度和训练过程的稳定性；在常规BP神经网络的基础上加入正则化算法，防止网络训练的过拟合。在选择网络层节点参数的细节上，简化输出层节点数，减少网络的训练参数。最后通过对印刷体字符样本进行识别实验，识别准确率达到98.97%，证明该方法可有效地提高字符的识别率。

Abstract: In order to solve the problem that the conventional BP neural network algorithm in character recognition algorithm is easy to cause over-fitting and leads to training failure, the character recognition rate is not high, and the effect of traditional template matching method on character recognition is not good, an improved BP neural network character recognition algorithm based on PSO is proposed. Particle swarm optimization algorithm (PSO) is used to calculate the initial weight and offset of BP neural network, which improves the convergence speed of network and the stability of training process. Regularization algorithm is added based on conventional BP neural network to prevent over-fitting of network training. In the details of selecting the node parameters of the network layer, the number of nodes in the output layer is simplified and the training parameters of the network are reduced. Finally, through the recognition experiment on the printed character samples, the recognition accuracy reaches 98.97%, which proves that this method can effectively improve the character recognition rate.