Train wheel tread defects detection based on image registration

Abstract

There are few feature points on the wheel treads, and the registration is between two different objects, wheel tread registration cannot be solved by traditional method such as key points matching. A system is built to detect the defects of tread. We design the algorithm used in this system. Firstly, we use Helmholtz Principle to detect lines and ellipses in images. Then the edges of the tread are located by magnifying and translating the easily-detected ellipse, registrations are performed between template and under-detecting tread image. At last, we locate the distinctions between the two images, which are the candidates of the tread defects. Experimental results indicate that the precision of segmentation and registration of treads can satisfy the system's requirements of rapid-speed and stabelization.