Abnormal Target Recognition For D-series High-speed Train Body Images Based On Convolutional Neural Network

D-series high-speed trains are an important means of transportation in China.In recent years,high-speed trains continue to speed up and become popular,so the safety concerns have been raised.Abnormal target recognition for high-speed trains is a crucial part when trains are running.Currently,abnormal targets are still mainly examined by inspectors.By this way,trains should stop,so that instant detection is not available and the labor cost is high.In this thesis,computer vision technology is adopted to collect and process the train images to achieve abnormal target recognition automatically and intelligently.Owing to no parking,this method has high detection efficiency and strong real-time performance.With the assistance of this system,the workload of inspectors can be greatly reduced.First,the high-speed train image abnormal target recognition system is designed according to the characteristics of the high-speed trains whose bodies are very long and its software is programed based on MFC.This system can complete a series of operations such as train body catenating and alignment,and by computing the structural similarity between the current train body image and the standard one that has no malfunction,the differences between them are detected.These differences are potentially dangerous targets for the current train.However,most of the differences are caused by false alarms such as stains,muds and luminance differences.Because these differences are not dangerous,they should be removed from potentially dangerous targets.Considering that the aforementioned differences have no stable feature and the feature extraction for train images are difficult,traditional image processing and pattern recognition method cannot satisfy the robustness and universality of the algorithm.Thus,deep learning is introduced to the abnormal target recognition system.Different to the traditional classification tasks,the input of difference classification network is two images,namely,the baseline image and the current one.The neural network should first detect the difference between the two images and further identify whether it is potentially dangerous.To integrate the two-image information,two convolutional neural networks are presented to implement difference classification.To improve the performance,we introduce a multi-shape training strategy.Extensive experiments demonstrate that this strategy can improve the score of each evaluation index of the difference classification network by a large margin.Difference classification is not limited to high-speed train;it is a general algorithm that can be easily transplanted into other fields in which decision should be made with two states.The key components of high-speed train are important.However,owing to the lack of malfunction images,it is impossible to utilize deep learning to classify.Considering that the abnormality of key components is mainly caused by bolt losing,so in this thesis,the key component abnormal target recognition task is converted to bolt detection.Moreover,there are many bolts on other parts of the train.Therefore,the generalization of the network can be improved by learning how to detect the bolts of other parts and the performance of key component bolt detection can finally be improved.