Research On Rail Crack Detection Algorithm Based On Convolutional Neural Network

Track detection is an important means to maintain the safe operation of the railroad,in recent years the construction of China’s railroad has also entered a period of rapid development,with the railroad speed,heavy load and high density operation,makes the rail surface easily crack structural damage,if not timely to crack detection maintenance,will bring great threat to the operation of the railway safety.In order to ensure the safety of railway operation,it is necessary to detect and identify rail cracks in a timely and accurate manner.In this thesis,a rail surface crack detection and recognition method based on convolutional neural network is proposed.In view of the shortcomings of the existing model methods in the practical application process,the model is improved based on the adaptability of rail crack target detection to meet the requirements of efficient and fast recognition.The main research contents of this thesis are as follows:(1)Review the literature at home and abroad,understand the research progress in related fields at home and abroad,and introduce the relevant knowledge of convolutional neural networks.In terms of target detection,three target detection methods in different directions are introduced,among which the target detection method based on deep learning is mainly introduced,and two detection algorithm models in different stages are introduced in detail.This thesis studies the YOLOV5 s algorithm based on V6.1 version,and compares it with other detection algorithms to analyze its advantages,which lays a foundation for the later research.(2)In the preliminary preparation work,the data set containing the rail crack target is first collected,and then the crack label is made for the original image data.According to the requirements of the experimental training for the data set,different data enhancement methods such as splicing and flipping are used to enrich the data set and enhance the fitting of the network training.Then,the script format is converted by the code program to make it conform to the data format of the convolutional neural network training secondly,the clustering algorithm of the anchor frame in the network model is improved.It is found through experiments that the clustering results of the anchor frame have a great influence on the detection efficiency of the target.In this thesis,the K-means++ algorithm for adaptive anchor frame calculation and scaling transformation is compared.The superiority of the improved algorithm is verified by clustering center diagram and experimental results.(3)Aiming at the problem that the current crack recognition network structure is complex and the network depth is deep,it cannot be applied in the scene with high real-time requirements.An improved method based on lightweight network is proposed.The lightweight convolution method GSconv is used to replace the ordinary convolution method.The VOV-GSCSP module that can retain the information interaction between channels is introduced and applied to the neck network according to the convolution characteristics.The accuracy of detection is maintained while reducing the complexity of the model.Experiments on the rail crack data set show the superiority of the improved strategy.Aiming at the problem of poor detection effect of small targets,in order to maximize the utilization of shallow information,a cross-layer fusion structure is added to the feature pyramid network,and the information in the backbone network is integrated into the down sampling process of the path aggregation network.The feature information of the high and low layers is fully integrated to improve the accuracy of target detection.The improved network model is compared and analyzed on the data set.The experimental results show that the improved YOLOV5 s algorithm obtains an average accuracy of 57.5%,which is higher than the average detection accuracy of the improved network model.For the problem of poor detection effect of small cracks in rail cracks,false detection and missed detection,the attention mechanism is introduced with reference to the current mainstream methods.The Coordinate Attention mechanism and Transformer structure are added to the end of the feature extraction network to strengthen the extraction of deep semantic information,so that it retains the small target features of rail cracks to a greater extent.The experimental results on the rail crack dataset show that the algorithm with attention mechanism has excellent detection performance.