基于改进卷积神经网络的铁轨伤损图像识别

铁轨探伤技术的可靠性关系到铁路运行的安全性。分析BP神经网络、卷积神经网络算法在图片识别中的优势，提出一种结合BP、卷积网络的新算法应用于铁轨伤损检测。改进算法利用卷积神经网络对铁轨样本进行特征提取，仅一次前向运算获得低维度铁轨图，再由BP神经网络对低维度铁轨图特征进行分类训练与测试。实验结果表明，改进算法在已训练好的模型测试中得到较好的误差收敛曲线与较高的测试精度，与BP算法、卷积算法相比，该算法训练时间更少，对铁轨伤损图片识别效果更好，在铁轨伤损检测方面有较好的应用前景。

Recognition of Track Damage Images Based on Improved Convolution Neural Network

The reliability of rail flaw detection technology is related to the safety of railway operation.The advantages of BP neural network and convolution neural network (CNN) algorithm in image recognition are analyzed,and a new algorithm combining with BP and convolution network is applied to detect the damage of railway tracks.The CNN is used to extract the feature of the rail sample,and the low dimension rail map is obtained by only one forward operation.Then the BP neural network is used to train and test the characteristics of low-dimensional track.The experimental results show that the improved algorithm obtains better error convergence curve and higher test precision in the trained model test.Compared with BP algorithm and convolution algorithm,the algorithm has less training time and better recognition of rail damage images,so has a good prospect in the field of rail flaw detection.