Pavement crack detection using hessian structure propagation

Abstract

Pavement images are widely used in transportation agencies to detect cracks accurately so that the best proper plans of maintenance and rehabilitation could be made. Although crack in a pavement image is perceived because the intensity of crack pixels contrasts with that of the pavement background, there are still challenges in distinguishing cracks from complex textures, heavy noise, and interference. Unlike the intensity or the first-order edge feature of crack, this paper proposes the second-order directional derivative to characterize the directional valley-like structure of crack. The multi-scale Hessian structure is first proposed to analytically adapt to the direction and valley of cracking in the Gaussian scale space. The crack structure field is then proposed to mimic the curvilinear propagation of crack in the local area, which is iteratively applied at every point of the crack curve to infer the crack structure at the gaps and intersections. Finally, the most salient centerline of the crack within its curvilinear buffer is exactly located with non-maximum suppression along the perpendicular direction of crack. The experiments on large numbers of images of various crack types and with diverse conditions of noise, illumination and interference demonstrate the proposed method can detect pavement cracks well with an average Precision, Recall and F-measure of 92.4%, 88.4%, and 90.4% respectively. Also, the proposed method achieves the best performance of crack detection on the benchmark datasets among methods that also require no training and publicly offer the detection results for every image.