Research On Image Recognition Technology For Track Bolt Fastener Based On Multi Feature Fusion

With the increasing mileage of railway operation,the importance of safety inspection in rail system is becoming more and more important.Track bolts fastener as an important part of track system is regularly inspected for its integrity.In order to change the present situation of traditional artificial visual along the railway which is waste time and money,using image processing method of fastener non-contact integrity detection gradually become an important means.In order to improve the detection efficiency and detection speed of the fastener image detection process.In this thesis,the preprocessing,fastener location,feature extraction and model classification are discussed and studied in the process of fastener detection.First of all,in the image preprocessing stage,in order to solve the fastener image light too strong or too dark,the image using the histogram equalization method to adjust the brightness,the median filtering method to remove image noise,effectively improve the quality of the image,and the effects were verified.In the positioning stage of fastener images,a new projection method based on gray projection integration is applied to locate the image,which effectively overcomes the problem that the straight line features can not be correctly identified when the image is deflected.The single features usually have certain limitations,in dealing with the complex conditions of the picture,often only for a class of images have better effect,the integration of MB-LBP and PHOG feature fusion,feature generation has two kinds of characteristics,and more able to adapt to the shooting environment,image feature recognition under complex conditions and requirements.Finally,in order to solve the fusion feature dimension increase training speed slow down,introduction of the Adaboost-SVM hybrid classifier,after reaching a certain set of conditions the classification efficiency of Adaboost classifier will be reduced,and then the remaining samples into SVM classifier.The simulation test,the algorithm effect by using Matlab software.The results show that compared with the single feature detection methods and other types of image feature fusion in have certain effect to improve the accuracy,at the same time using Adaboost-SVM hybrid classifier after the training speed has been significantly improved.