Study On The Method Of Railway Fastener Image Feature Extration Ang Recognition

With the leap-forward development of the railway in our country, many new challenges appear in the railway transport safety guarantee system. As one of the main infrastructures, railway fastener can alleviate the mechanical vibration and shock when the train passes over the orbit by its elastic deformation and absorbing energy, however, it may lose or partly break because of that the fastener often tolerate the periodic bending and torsion alternating stress. In addition, the fatigue fracture or missing of the fastener become more and more serious, with the continuous development of the high speed railway and heavy haul freight train. The fastener is so important role in railway transportation that the monitoring and detection of its working conditions have a vital significance for protecting the railway transport safety. In order to meet the need of the railway development, the designing of fastener automatic detection system which is high information sharing and reliability has become the emphasis of the railway transport safety system. As a concerned frontier topic and a new application direction, the automatic detection of the railway fastener is realized by the image analysis and processing based on the computer vision. Inspired by previous research at home or abroad, the key techniques of fastener image detection are mainly researched in this paper, including the accurate location and segmentation of fastener sub-image, the effective extraction of image feature, and the image classification and recognition algorithm. A strong robustness fastener detection method is put forward in this paper. The main contributions are as follows:First, due to the fastener only occupy a tiny area in the original image, it is time-consuming and is difficult to get the ideal detection result if we detect the fastener directly in the original image. To avoid the disturbance caused by the other objects except for the fastener in the image, an image segmentation algorithm which including rough and accurate location is proposed. First of all, the average gray value and significant difference between rail, sleeper and railway ballast in the orbital images are calculated, respectively, and then a symbol rate function based on two kinds of different information is build for getting the position of the rail and sleeper, and we get rough fastener area position by crisscross localization based on the position of the rail and sleeper. Then, the binary image corresponding to rough segmented fastener image is obtained according to the gray invariance between the sleeper shoulder and the image background, and we get the position of the sleeper shoulder by the projection method based on the linear feature of it in the binary image, after that, the accurate fastener position is obtained according to the positional relationship between the sleeper shoulder and fastener.Secondly, the effective EAHOG-MSLBP fusion feature is put forward to describe the characteristics of the fastener image. First, a fastener EA edge awareness method based on template is designed, and an improved EA-HOG edge gradient feature description algorithm based on the fastener edge is presented, which is an approximation of the fastener shape feature and it is not sensitive to light and color change. Then, to overcome the weakness that the traditional binary pattern and its related algorithm can only express the micro texture pattern and can't catch the structural differences between the fastener and image background, a macroscopic Local binary pattern (MSLBP) is proposed to extract the macro texture of the fastener in the image. Finally, the hierarchical weighted fusion algorithm is used to combine the EA-HOG edge gradient features and the MSLBP texture features into EAHOG-MSLBP fusion feature.Finally, for solving the problem that the partially broken fastener is difficult to recognition and classification, two virtual sample are obtained from the original fastener image by image symmetry operation, and the direct detection of the original image is turned into the indirect detection of two virtual symmetry samples, this algorithm breaks through the limitation of the direct detection by traditional detection method on the basis of computer vision. In addition, a weight accumulation sparse representation classification recognition algorithm is designed for the detection of two virtual fastener samples. The first step of this algorithm is to explore the expression ability of the all training samples to the test sample, and selectively reserve K training samples in each class of training samples which have stronger ability to express the test sample; the reserve K training samples in each class are used to linear express the test sample in the second step of this algorithm, and the K coefficient values in each class are accumulated and viewed as the contribution ability for the detection of the test sample, finally, the test sample is classified to the class that has the largest contribution ability.Finally, The conclusion of the paper summarizes the full work in this paper, and some research issues about the fastener detection theory and application of the fastener automatic detection technology it also pointed out in further study.