Railway Fastener Detection Using Gaussian Mixture Part Model

Herein, the shape deformations in the collected images of railway fasteners, the large illumination difference in the fastener image, and the partial occlusion of the fastener by foreign objects is addressed. A Gaussian mixture part model (GMPM) algorithm is proposed to address these issues. This model is based on previous research of deformable part model; however, it was further refined by combining with a Gaussian mixture model (GMM) algorithm. The image gradient was calculated by combining the edge characteristics of the fastener image and application of the improved Roberts operator to the image. Normalized histogram of oriented gradient (HOG) features were used as the basis of the GMPM algorithm. The image was divided into a number of regions considering the shape of the fastener, and a star connection was used to measure the relative positions of the various subsections of the subdivided image. Cosine similarities were computed to measure the similarity of the HOG features of the different parts of the image. The part model was solved iteratively by using the GMM with an expectation-maximization algorithm. By using the GMPM algorithm to detect defects in railway fasteners, an average accuracy of 90.27% for detection rate is achieved, while the average missed detection rate, and the average false detection rate are 3.16% and 9.80%, respectively.