Surface Crack Detection In Ballastless Slab Track Of High-speed Railway Based On RetinaNet

In recent years,China’s demand for railway transportation is increasing.As a symbol of railway development,high-speed railway plays an irreplaceable role in a country and develops rapidly.However,under the high bearing capacity and frequency of long usage,ballastless slab track of high-speed railway will be subjected to continuous stamping and appears various degrees of surface crack,which will reduce the load capacity and smoothness of the track then affect the safety of train operation.Therefore,surface crack detection in ballastless slab track of high-speed railway has become an important work to ensure the safe operation of high-speed railway.At present,the technology of deep learning develops rapidly,possesses the characteristics of high detection accuracy and better recognition effect,it is widely used in object detection.Therefore,method of surface crack detection in ballastless slab track of high-speed railway based on deep learning can recognize surface cracks accurately and quickly to prevent accidents.In view of the problems of small sample quantity,large difference between scales and unbalanced crack category on the surface cracks in ballastless slab track of high-speed railway,the following work is mainly carried out in this paper:(1)In order to solve the problem in small sample,improve generalization ability of the network model and avoid the underfitting phenomenon due to the lack of sample training,the data set is enhanced by horizontal flipping,vertical flipping,rotation,scaling,clipping and adding random noise.(2)RetinaNet is selected as the basic detection network to solve the problem of large scale difference of surface cracks in ballastless slab track of high-speed railway and a multi-level feature pyramid network is used to fuse the information about the characteristics of different levels extracted from the backbone network,so as to fully express the image feature information.Meanwhile,in order to improve the classification and positioning accuracy of surface cracks of different scales in the detection network,adaptive anchor optimization is used to solve the problem of mismatching between the classification and positioning confidence of surface cracks in the detection process.Furthermore,for the problem of unbalanced sample category,class-balanced weight items are introduced into focal loss function to make the sample of each category rebalance the loss,which alleviates the impact of class imbalance on detection performance.(3)To further improve the detection speed of surface cracks in ballastless slab track of high-speed railway,lightweight network of MobileNet is used as the feature extraction network.To ensure certain detection accuracy,multiple parallel atrous convolution and spatial pyramid pooling network are added to the backbone network.Among it,parallel atrous convolution is used to construct receptive fields of different sizes to detect multi-scale surface cracks and the module of spatial pyramid pooling network is done to ensure the output size of feature map to avoid the loss of the contextual feature information of multi-scale surface cracks.In the feature fusion network,a new structure of feature pyramid network is constructed to fuse the depth and shallow features of surface crack between two adjacent layers in the backbone network and combines with the original feature pyramid network of RetinaNet to enhance the feature information.In brief,aiming at the existing surface cracks in ballastless slab track of high-speed railway,research and experiments based on detector of RetinaNet are proceeded respectively in this paper from aspects of data set construction,model building and optimization,which improves the accuracy and real-time detection of surface cracks in ballastless slab track of high-speed railway and qualify for certain theoretical and value of engineering application.