Research On Track Slab Crack Detection Technology Based On Deep Learning

This improvement of national economy has led to the growth of comprehensive strength and the speed of infrastructure construction in China.As the lifeline of the national economy,the railway becomes a key infrastructure and a major livelihood project.In order to ensure the personal property of people,it is necessary to detect the cracks of track slab timely and effectively.At present,CRTSII slab track has been laid all over this country as the mainstream ballastless track.Considering the operation time of rail transit,the image acquisition of track board can only be carried out at night,and the severe shooting condition leads to uneven illumination and low contrast in final imaging photos.It is difficult to detect the cracks in such photos.In order to accurately detect the cracks in track slab images,deep learning is used to improve this crack detection technology of track slab from two aspects:(1)In order to solve this problem of false detection and missed detection of track slab cracks at image level,improved basic convolution neural network model combined with voting mechanism.Based on attention mechanism,feature maps are grouped to form multiple groups of feature vectors,which are sent to the weak classifier respectively.These parameters that can be learnable,and the decision information of all weak classifiers is gathered to form a strong classifier.This method could reach 97.1% accuracy in the track slab data set。(2)In order to further detect the cracks in track slab,it is necessary to determine the location information of cracks.However,in the fully convolutional neural network,using bilinear interpolation to expand resolution may only make the model achieve a suboptimal solution.Through this autoencoder training of crack label,the mapping relationship of reconstruction is obtained,which is used to replace the relatively simple bilinear interpolation method.Finally,the MIo U of segmented network can reach 64.8%,and the unmarked crack area even could be detected.Experiment shows that this improved algorithm can not only reduce parameters of this model,crack can still be accurately identified even in complex environment.