Research And Application Of Key Technologies Of Railway Image Intelligent Analysis Based On Deep Learning

Railway transport is an important part of China's integrated transport system,and in recent years it has made rapid development.Safety,as a prerequisite for railway operation,is the priority of railway operation and maintenance.In order to ensure the safe operation of the vast and complex railway network,China has established a safety inspection and monitoring system covering the mainstream professions.Several security management information systems have been established for infrastructure,mobile equipment,internal and external environments,etc.With the generation and accumulation of a large amount of detection images,there is an urgent need to carry out analysis and mining of massive images to enhance the safety of railway.Limited by the performance bottleneck of traditional image analysis techniques,which currently rely mainly on manual analysis.On the one hand,the large amount of detection iamges places a heavy workload on the professional,resulting in a significant drain on human and material resources.On the other hand,the accuracy and coverage of image analysis is directly related to the capacity and experience of professional,which makes it difficult to ensure the quality of the analysis work.In addition,since manual analysis is generally performed after image acquisition,there is no way to ensure the timeliness of the analysis.With the development of artificial intelligence,deep learning has made breakthroughs in computer vision.Therefore,it is extremely necessary to apply deep learning methods to solve the problem of intelligent analysis of railway detection images.The railway detection images cover many scenarios such as key component,rail and tunnel detection.The characteristics are as follows,such as a variety of resolution scales,image quality and shooting angles,etc.In order to promote the application of advanced technologies such as artificial intelligence,big data and Internet of Things in the field of railway,this thesis has conducted a series of studies.The main work and results are as follows.(1)Based on residual network,a scene classification method that integrates transfer learning and model visualization is proposed to solve the task of railway scene classification.This method addresses the problem that the small amount of railway scene data makes model difficult to converge,and experiments conducted on the Railway12 dataset validate the effect.Based on the idea of transfer learning,the network that have been trained on large scale dataset is fixed with their shallow parameters,and the deep parameters of the network are funtuned on the railway image scene classification dataset.Based on model visualization approach,the interpretability of the railway scene classification model is enhanced.Experiments on the Railway12 dataset showed that the proposed method can achieve classification accuracy of 82.6%(Top1)and 95.3%(Top3).At the same time,the visualization method helps to reduce the disequilibrium between classes.Experiment proves that the visualization method can improve the classification accuracy by about 2%.(2)Based on Faster R-CNN,this thesis proposes a two-channel defect detection method based on supervised learning theory.The validity of this method are verified by experiments performed on the railway key component defect detection dataset and NEU defect detection dataset.The proposed method divides the defect detection task into component detection and defect classification channels.The improved Faster R-CNN method is used to improve key component detection performance.In addition,SRGAN and RAISR are combined to achieve super-resolution improvement for small component.Finally,a separately trained defect classification network is used to determine whether componnet contains defects.Experiment conducted on images collectd by TEDS demonstrated that the mAP value can reach 0.792,which is a significant performance improvement over other object detection methods.(3)Based on semi-supervised learning,a multi-mapping anomaly detection method is proposed to solve the problem of data imbalance.Experiments conducted on MNIST,CIFAR10 and railway dataset proved the effectiveness of this method.The anomaly detection method proposed in this thesis is based on the BEGAN to generate normal samples,and learn the multi-mapping pattern of normal samples.During the testing stage,positive and negative samples are distinguished by defining an appropriate threshold.Experiments performed on images collected by the 4C detection system proved that this method has a significant performance improvement over existing semi-supervised learning anomaly detection algorithms and can be applied in real-world anomaly detection tasks.(4)Based on the Tensorflow-Serving model deployment mechanism,we propose a deep learning deployment scheme that fuses cloud and edge to solve the problem of high transmission bandwidth pressure on the existing method networks.The validity of this method is verified by the construction of a railway image big data intelligent analysis platform.Edge computing has the advantages of saving bandwidth,storage costs and enabling real-time response.Model lightweighting techniques such as quantization,pruning,and weight sharing are used to build an edge-end network model suitable for small arithmetic scenarios,which enables real-time analysis for detection video.The cloud center can realize distributed training and inference of complex models.By fusing historical data from multiple terminal devices for analysis,business forecasting and early warning can be achieved.