Research On Classification Of Rail Damage Images Based On Deep Learning

Railway plays an important role in the transportation of China.Therefore China is extremely concerned on railway maintenance and transportation safety.What's more,the detection and classification of rail damage is a crucial part in railway maintenance,which is the reference to take corresponding maintenance measures.Image classification for the severity of rail damage is the target problem that this paper needs to solve.Traditional image classification methods are generally carried out by the steps of extracting features,reshaping features and classifying features.In this paper,extracting features method contains extracting HOG feature,SIFT feature and LBP feature.Reshaping features method uses the combination of the bag-of-word model and the K-means clustering algorithm.And the classifier is support vector machine.At present,with the deepening of deep learning research,the accuracy of deep learning has exceeded the traditional methods in ImageNet Large Scale Visual Recognition Challenge,which inspires us whether we can apply deep learning to classify rail damage image.For the classification task of rail damage images,this paper proposes an image classification method based on deep learning.At first,data augmentation technology which contains rotation flipping,grayscale change,adding noise method is used to augment the quantity of original dataset images as ten times as before.And then the order of the rail damage images is messed up to make the TFRecord data file.Secondly the idea of building neural network is summarized,and the VGGNet network is modified to adapt for rail damage classification.The new network is named RDC-VGGNet,and the subsequent experiment shows that RDC-VGGNet runs 12% faster than VGGNet.Finally,this paper shows training steps and training parameters of RDC-VGGNet on TensorFlow.The visualization tool TensorBoard is used to monitor related parameters and a network model that achieves the expected results has been trained.Three traditional methods: HOG+SVM,SIFT+SVM,LBP+SVM,are selected to compare with RDC-VGGNet to test the classification effect of rail damage images.Experiments are designed and indicators such as TOP-1 accuracy rate,precision rate,recall rate,ROC curve,running time,are established to evaluate the performance of each model.Experimental results show that the TOP-1 accuracy rate of VGGNet is 93.04%,the precision rate is 95.61%,and the recall rate is 94.53%,which is higher than the traditional methods.In addition,the ROC curve and running time of RDC-VGGNet are also superior to the traditional method.Above experimental results highlight the superiority of deep learning in image classification tasks.