3D Pavement Crack Image Detection Based on Anisotropy Measure

In order to detect cracks precisely and completely, a novel automatic cracking detection algorithm based on 1 mm/pixel 3D pavement images was proposed with three core parts: anisotropy measure calculation and self-adaptive threshold optimization, depth verification and multi-resolution denoising processing. First of all, considering pixel features of a pavement image, anisotropy measure was calculated for each pixel based on mean values and standard errors of four linear neighborhoods at 0 degrees, 45 degrees, 90 degrees and 135 degrees to represent significance of orientation, and an optimal anisotropy threshold was determined by the OTSU method to segment the pavement image into two parts, that is, strong orientation and weak orientation pixels. Then, a depth threshold was set based on mean depth value in a square neighborhood with a radius of d, thus a potential crack image was formed by pixels with strong orientation and depth not bigger than the depth threshold. At last, the crack image was divided into several blocks and denoised by a moving window denoising method based on newly designed templates, from which final crack image was obtained. Algorithm tests were conducted based on 166 three dimensions pavement images(2 048×2 048) with various types of cracks. The result shows that the proposed algorithm achieves relatively high precision (averaging 91.57%) and recall (averaging 81.29%), leading to an mean F1 score of 84.26%, thus outperforming seeds based approach(mean F1 score: 69.19%), Canny edge detection(mean F1 score: 8.15%) and OTSU segmentation(mean F1 score: 5.11%).