Recognition and evaluation of bridge cracks with modified active contour model and greedy search-based support vector machine

Concrete cracks are the most important representation for evaluating the bridge health condition and conducting to take appropriate actions to optimize expenditure on maintenance and rehabilitation. In this paper, we develop a fully-automatic machine learning based algorithm for extracting cracks from concrete bridge images, which combines a modified region-based active contour model for image segmentation and the linear support vector machine using greedy search strategy for noise elimination. In practice, the crack detection is a challenging problem because of (1) subtle difference between the cracks and the noises, (2) inconsistent intensity along the cracks, and (3) possible shadow regions with similar intensity to the cracks. To solve these problems, the proposed method consists of three steps. First, we build a high-precision image acquisition framework, which can automatically collect image sequences from the lower bridge slab and fuse the multiple sensor data for computing crack parameters. Second, we develop a modified region-based active contour model combined with the iterated Canny operator for the concrete image segmentation. Finally, we utilize the novel feature selection approach based on the linear support vector machine with a greedy search strategy for noise elimination. After that, we provide a crack width calculation method which combined the binary image with the gray scale image information. We evaluate the proposed method on a collection of 1200 real bridge images, which gathered from 10 existing bridges on various weathers, and the experimental results show that the proposed method achieves a better performance than several up-to-date algorithms.