Study On Enhancement And Extraction Of Low-SNR Pavement Cracks

Cracks are the most common distresses on the road pavement. Fixing a crack beforeits deterioration can greatly reduce the cost of pavement maintenance. Most traditionalimage-based approaches for pavement crack detection implicitly assume that pavementcracks in images are with high contrast and good continuity. However, this assumptiondoes not hold in practice due to1) uneven illuminance caused by pavement shadows,2)speckle noise brought by grain-like texture of pavement,3) low contrast between cracksand the surrounding pavement and intensity inhomogeneity along the cracks caused bycrack degradation and4) bad continuity of cracks incurred by the high sensitivity ofcrack-imaging results to the direction of illumination. In the above conditions, pavementcracks show their linear structures in a low-SNR manner, which brings great challengesto the automatic detection of pavement cracks. To address these problems, this paperconducts the following research.(1) Pavement shadows not only create uneven illuminance for pavement images, butalso undermine the intensity homogeneity of pavement cracks, which greatly increasethe difculty of identifcation of pavement cracks. To locate and remove pavement shad-ows is critical to the detection of pavement cracks. Considering that pavement shadowstypically hold a big penumbra area, we propose an intensity-based geodesic model tolocate the shadow area, as well as its penumbra area. Moreover, since traditional ad-ditive illuminance-compensation algorithm can not balance the texture detail betweenthe shadow area and the non-shadow area, we propose a multiplicative illuminance-compensation algorithm, which can improve the contrast of the shadow area to the levelof the non-shadow area by adjusting the variance. Then, a novel shadow removal al-gorihtm, i.e., GSR, is formed by integration of the above two components. GSR canautomatically locate the shadow and balance both the intensity and texture betweenthe shadow area and non-shadow area, and meanwhile preserve the cracks.(2) Due to the particle texture of pavement materials, the binary pavement imagesoften contain a lot of speckle noise, which results in low SNR of pavement cracks againstthe pavement background. Note that the tensor is ft for describing the point target,and the Gestalt laws embedded into the voting process can potentially infer out the linear saliency. We exploit a tensor-voting-based method for crack enhancement, whichfrst uses a ball voting to form the orientation at each token, and then applies a stickvoting to softly connect the neighboring tokens, and at last, conduct an eigen-featureanalysis to extract the linear saliency.(3) Since pavement cracks constantly sufer from the rolling of loaded wheels andthe weathering, crack degradation often exists and hence makes a low contrast betweenthe cracks and the pavement background, a bad continuity of the cracks as well. Toenhance these cracks, we exploit a method based on minimum-cost-path searching. First,we present a multi-scale F*algorithm for crack tracking in pavement images, which ismuch efcient than the original F*algorithm. Based on that, we propose an F*seed-growing algorithm, i.e., FoS, which achieves automatic selection of the tracking startand the tracking end. We develop an algorithm to automatically collect the crack seedsand subsequently apply a FoS process to enhance the cracks. We also experimentallystudy how the radius of seed growing impacts the enhancement results.(4) To extract linear structures from a set of spatial points, we propose target-point minimum spanning tree, i.e., T-MST. T-MST inputs the target points into agraph model, and then compute the minimum spanning tree by considering the Gestaltlaw of proximity. Since a crack is a linear structure in a macro perspective, T-MSTembeds the Gestalt law of continuity into a priming algorithm and extract the fnal linecurves. Based on T-MST, we proposed the FoSA approach and CrackTree approachfor pavement crack extraction. The former acquires the target crack points by an F*seed-growing algorithm, while the later adopts a sampling strategy to collect the targetpoints. A range of experiments demonstrate that FoSA is efective and efcient inextracting complex cracks featured with low contrast and bad continuity. Meanwhile,experiments on large-scale dataset show that the proposed CrackTree achieves a muchbetter performance than several existing methods.