Research On Abnormal Target Detection Method In Real-time Surveillance Images Of Electric Multiple Unit

Since the beginning of the 21 st century,China has begun to develop high-speed trains.High-speed trains play an increasingly important role in the national economy,its safety is highly important.However,due to its high speed and complex system,high-speed trains are prone to various abnormalities.Thus,the safety inspection of high-speed trains has evolved as a key technology.Trouble of moving freight car detection system(TFDS)is the most commonly used system for anomaly detection of high-speed trains.The TFDS first obtains images of the bottom and side of the EMU through multiple industrial cameras,and then the indoor inspectors identify and locate the abnormal targets in the image,which has low efficiency and couldn’t realize real-time detection.Image recognition methods can effectively solve this problem,however,existed methods are either have low accuracy or poor robustness.In this dissertation,we take eight abnormal targets in real-time surveillance images of EMU as examples,study the problem of abnormal target detection in real-time surveillance images of EMU,and aim to build a detection model with high accuracy and strong robustness.This research is highly significant to improve the TFDS and promote the researches on image-based theory and method for abnormal target detection.Besides,the proposed model could also be used to detect other abnormal objects with line shape and patch shape.The main contents are as follows:(1)To improve the accuracy and robustness of existed image recognition methods for abnormal target inspection in real-time surveillance images of EMU,the deep convolutional neural network and salient object segmentation technology are studied,and a regional convolutional neural network based abnormal target detection model(RCNN-ATD)is constructed.The RCNN-ATD model divides the task of abnormal target detection in the real-time surveillance images of EMU into three sub-tasks: feature extraction,abnormal target classification and location,and abnormal target segmentation.These sub-tasks are optimized and solved respectively to reduce the learning difficulty.In RCNN-ATD,a simple and effective image preprocessing method is proposed to enrich the feature information contained in the samples,and the deep convolutional neural network is used for extracting image features and establishing the feature descriptor of abnormal targets in the image.Then,it constructs a classification and regression module to classify and locate the abnormal target in the image.Finally,it proposes a convolution-deconvolution based abnormal target segmentation method(C-D-ATS)to segment abnormal targets in images.Due to small dataset,the transfer learning method is used in the training process.The model is first pre-trained using COCO dataset to fully utilize the feature extraction ability of DCNN.The RCNN-ATD model was used to detect five abnormal targets in the real-time surveillance images of EMU,including oil leakage,shaft scratch,foreign object,end cover anti-shedding chain breaking and anti-loose wire breaking.Experimental results verified the effectiveness of RCNN-ATD’s architecture.(2)To further improve the accuracy of RCNN-ATD model in inspecting multi-scale abnormal targets,image pyramid and multi-scale feature representation methods are studied.A Gaussian pyramid feature fusion(GP-FF)method is proposed to construct a multi-scale feature representation of the image,the RCNN-ATD model is improved and a multi-scale feature fusion based abnormal target detection model(MS-ATD)is constructed.The MS-ATD model was used to detect five abnormal targets in real-time surveillance images of EMU,such as oil leakage,shaft scratch,foreign object,end cover anti-chain breaking and anti-loose wire breaking.Experimental results showed that the MS-ATD model with GP-FF method achieved higher accuracy than RCNN-ATD model.(3)To further improve the abnormal target segmentation precision,the image contour extraction methods are studied.An abnormal target segmentation(ATS)method is proposed,the MS-ATD model is improved and a multi-scale feature fusion based highly precise abnormal target detection model(MS-HPATD)is constructed.The MS-HPATD model was used to detect five abnormal targets in real-time surveillance images of EMU,including oil leakage,shaft scratch,foreign object,end cover anti-chain breaking and anti-loose wire breaking.The results showed that the MS-HPATD model with ATS was more precise than MS-ATD model in contour detection.(4)To further improve the robustness of the model,the confrontation training is studied.The confrontation sample construction methods and the model training methods are reviewed.A confrontation sample based robust enhancement method(CS-REM)is proposed to improve the robustness of the proposed model.(5)For the detection of small-scale abnormal target of EMU,the characteristics of small-scale EMU abnormal targets are studied and an improved MS-HPATD model(Improved MS-HPATD)is constructed.Specifically,a new data augmentation method is proposed to increase the learning probability of small-scale abnormal targets,and an anchor box design method is proposed to optimize the parameters of MS-HPATD to improve the detection probability of EMU small-scale abnormal targets.The improved MS-HPATD model was applied to the detection of pin missing,bolt missing and rivet missing.The results showed that it could accurately detect these three small-scale anomalies,and its detection accuracy was higher than that of the state-of-the-art image recognition methods for small-scale abnormal target detection.