Road Surface Crack Detection: Improved Segmentation with Pixel-based Refinement

Cracks are among the most commonly found road surface degradations, requiring periodical road surveys for monitoring pavement quality. Images of road pavement surface can be automatically processed, typically employing segmentation algorithms to identify cracks. However, a set of distinct connected components often result, leading to the detection of several independent crack segments, although they may belong to the same pavement surface defect. This is often observed for cracks that exhibit a longer linear development or present several branches. This paper presents a new strategy to identify cracks on images captured during road pavement surveys, even when those cracks appear with a complex shape. The proposed crack segmentation algorithm includes two stages: (i) selection of prominent "crack seeds", adopting an efficient segmentation procedure, after appropriate image smoothing, minimizing the detection of false positives; (ii) iterative binary pixel classification, into the crack or non-crack classes, extending the "seeds" to identify the complete crack shape. The paper also tests the combination of the proposed two stage crack segmentation with three smoothing techniques, to evaluate their suitability for crack detection. As a final step the system classifies the identified cracks as longitudinal, transversal or miscellaneous types. Tests performed with images acquired from different types of sensors (active and non-active), show improved crack segmentation results.