Train Bottom Bolts Image Recognition Technique Based On SIFT

The vehicle abnormality detection plays an important role in train test. Many unsafe factors exist in the bottom of train, particularly highlighted the importance of bolts. On the past, the staff enter the bottom of vehicle to detect by naked eye for conventional, but in such a way, it is not only time-consuming, but also a waste of manpower. So it is necessary to design a reasonable and effective vehicle anomaly detection scheme. In current, the technology of industrial linear CCD(Charge-Coupled Device) camera capture images has been relatively mature, and digital image processing methods are also emerging, it is visible feasibility display to achieve the identification and detection of abnormal vehicle bolts based on train image processing. Therefore, this thesis does research of image feature extraction operators suitable for train based on the above research background and significance. After comparative and analysis, it is confirmed that using improved SIFT (Scale-Invariant Feature Transform) algorithm combined with neural network algorithm to achieve the purpose of train bolts identification.Regarding the characteristics of the train bolts image got by industrial CCD cameras: local invariance. Prior to the survey of feature extraction operator, and ultimately determine to using SIFT feature extraction operator because of its rotation, scaling, partial invariance. Subsequently, the thesis focuses on the image recognition algorithms of train abnormal bolts based on SIFT feature extraction operator.To extract the initial features by SIFT feature extraction operator, to get exact match feature points by RANSAC(RANdom Sample Consensus), to correct image rotation and scaling, to obtain and display fault marks by template matching, enabling detection abnormal of train bottom images.To extract feature points of positive and negative sample bolts through SIFT feature extraction operator, combined with repeated training of SVM(Support Vector Machine), to get the robustness of the strong classifier, enabling detection abnormal of train bottom bolts images.The thesis studys two kinds of train bolts abnormal image recognition algorithm based on SIFT feature extraction operator. It can automatically extract features’information in the image and establish good feature descriptors. It can automatically match the feature points by setting a threshold parameter, the error and duplicate matching feature points will be automatically removed meanwhile. The rotation angle and zoom ratio can also be obtained by the test image and the template image, then automatic calibration orientation and size of test image to the template image. It can locate abnormal area by SSD (Sum of Squares of Deviations) and NCC (Normalized Cross-Correlation), then establish positive and negative sample bolts image to obtain strong robustness classifier by SVM. Finally, it can achieve the identification and marking bolts.Experimental results show that train image recognition algorithm based on SIFT feature extraction operator have encountered a good fit for abnormal rotational, scaling, local invariance of vehicle images. And train bolts image recognition algorithm based on SIFT feature extraction operator combined with SVM classifier have encountered good recognition on train bolts with 94% recognition rate.The thesis studys two algorithms to extract image feature points quickly and accurately, established feature descriptors are completely, with short running time and low CPU occupation, at the meanwhile, error and duplicate matching feature points are removed efficiently, besides abnormality positioning accuracy, and bolts identification marked clear.