Abstract: Traditional BP neural network has several shortcomings that training speed is slow, probability of error is great and it is easily falling into local minimum. A kind of improved composite error function, which can replace the traditional global squared mean error function, is presented. This novel function can improve the learning rate of the network, and a new BP neural network is introduced which can adjust the learning rate hierarchically and dynamically to classify the cracks in road surface pictures. The experimental results show that, compared with traditional methods, the improved algorithm has achieved a dramatic improvement in precision and speed.

Key words: BP neural network, crack classification, composite error function, stratified dynamic adjustment

摘要： 针对传统BP神经网络训练速度慢，误差大且易陷入局部极小值的缺点，设计了一种改进的复合误差函数来代替传统的全局均方误差函数以提高其学习率，同时采用了改进的分层动态调整不同学习率的新BP神经网络对路面裂缝图片进行分类。实验结果表明，与传统方法相比，改进后的算法在检测精度和速度上有了明显的提高。

关键词: BP神经网络, 裂缝分类, 复合误差函数, 分层动态调整