Anisotropic clustering on surfaces for crack extraction

Machine vision provides an efficient way for the automatic crack detection of civil structures. However, it is still very challenging to extract the small cracks embedded in noisy background. Even some very recent methods require manual intervention or omission of crack width. In this paper, we aim at extracting such inconspicuous cracks automatically with width information preserved. The basic idea of the proposed method is to assign the pixel points to some arbitrarily shaped clusters, and then sift out the crack clusters according to their elongated shapes. Treating each gray-level image as a parametric surface, we devise an anisotropic clustering algorithm that exploits the geometric properties of the surface. By virtue of the geometric representation and the anisotropy, this algorithm solves the problem of separating adjacent objects while simultaneously grouping the fragments of a crack into the same cluster. Moreover, the globally convex segmentation model is incorporated into our method, serving as a tool that provides appropriate candidate points and important parameters for the clustering procedure. Experimental results on real images demonstrate that the cracks extracted by our method are very similar to manually traced ground truth cracks and thus can be used for measuring the widths of real cracks.