Surface Defect Detection Of Air-rail Based On Image Analysis

Suspended monorail(air rail traffic)is a new form of rail transit.In order to ensure the safe operation of the air rail traffic,surface defect detection is one of the very important part.The method for the detection of air-rail is little,that is,we only can know it from other surface defect detection.Therefore,we propose a method of surface defect detection based on image analysis.According to the technical demand of image analysis,we designed two detection methods: offline detection and online detection.The offline detection is the selection of rail surface defect detection algorithm of redundant dictionary based on sparse representation,the main idea is from a large number of defect free sample image feature extraction to construct dictionary and optimize the dictionary.Then,the sparsity of the sparse representation coefficient of the sample under the dictionary is obtained,and then the defect is determined according to the threshold set by an experiment.The accuracy and recall of this defect detection algorithm are high,but the time is not real-time,so it can be used for off-line detection of defects.Another method of on-line detection is selected as a quick detection method,it is based on a coarse detection and detailed detection,The coarse detection is to recover the positive samples as much as possible.The mean and variance of the images collected by the camera are calculated by using a sliding window of a certain size.A threshold is obtained and divided into two parts: the suspected defective samples(the defective samples and normal samples)and the normal samples.The second step is to use a generalized integrogram feature to deal with the two parts of the coarse detection,and to compress the features of the two parts by using the compression sensing algorithm.It is found that the normal part of the sample belonged to a Gaussian distribution,and then the calculated samples of the suspected defective portions are substituted for each dimension to compute the probability of meeting the Gaussian distribution to judge whether the sample is a normal sample or a defect sample.The STC tracking method is introduced in the experiment to track the defects detected in the previous frame to improve the speed of the experiment.After the defects are detected,the geometric features need to be calculated.The defects selected in the experiment are fineness,aspect ratio,and circularity,which are used to distinguish the two types of defects from scars and cracks.BP neural network is used to identify.In this paper,we used OpenCV and VS2013 to set up the experiment platform,and used SQLServer2005 as the background database to complete the whole experiment.The results of simulation and on-line experiment showed that the algorithm was effective and feasible.