As cracks on concrete bridges become severer and more frequent, methods of detecting cracks on concrete bridges have aroused great concern. Conventional methods, e.g., manual detection and equipment-aided detection, suffer from subjectivity and inefficiency, which increases demands for an accurate and efficient method to detect bridge cracks. To this end, we modify the existing percolation method and propose an enhanced percolation method, which detects cracks of concrete bridges automatically. The modification includes three improvements, which are (1) employing photo expansion to eliminate boundary effects, (2) varying shape factors to increase the accuracy of percolating unclear cracks, and (3) decreasing the number of neighbouring pixels to form candidate sets. Combined with the above three improvements, three versions of enhanced percolation methods utilizing three different shape factors are put forward. The numerical experiment on detecting cracks in 200 images of the bridge surface demonstrates the outperformance of the enhanced percolation method in precision, recall, F-1 score, and time compared with traditional detecting methods. The proposed method can be generalized on the application of detecting other types of bridge diseases, which is an advantage for designing, maintaining, and restoring infrastructures.
A broad spectrum of other approaches has been proposed for crack detection. Employing depth information of points in cracks and backgrounds is another way [36]. Cabaleiro et al. [37] presented an algorithm for the automatic detection of cracks in timber beams sampled by LiDAR data. They pointed cloud of the beam face, removed points inside the cracks, and projected them on a 2D coordinate system to identify crack outlines. Abdel-Qader et al. [38] used a PCA-based algorithm to detect bridge cracks. They compared the performance of methods using PCA alone, PCA with a linear feature process, and PCA on the local region. However, the accuracy of the three kinds of methods was affected dramatically by camera pose and distance from where images were taken. Yu et al. [39] used robots to collect bridge crack images, but this method required to label the start and end points in cracks manually. Oh et al. [10] used a crack tracking way to extract width and length information of cracks. Zou et al. [40] proposed the seed-growing approach for automatic crack-line detection through extending the algorithm, and based on this, they developed CrackTree [41], which is a fully automated method to detect cracks from pavement images.
caspio bridge download crack software
Some of the unclear crack pixels are ignored using the existing percolation method, especially in images where the colour difference between cracks and backgrounds is little. It is because the stride parameter takes a constant value. After a certain number of iterations, the threshold value T increases to a high level due to the contribution of . Then, even those background pixels with high pixel values are more likely to be included in the percolated area and change its shape factor from near 0 to 1. Therefore, the constant value of is replaced and set to be a product of the shape factor of the percolated area and the constant value [27]. Because the value of the shape factor after the first iteration is between 0 and 1, the increase of can be well controlled by the linear function. However, it is not the only way to realize the inhibiting effect; a quadratic function and a cubic function are designed to replace the constant value. These strides are called linear, quadratic, and cubic strides, and this strategy is called the slacken stride strategy. Their equations are as follows:where is a linear function of Fc, is a quadratic function of Fc, and is a cubic function of Fc. In equation (5), when the shape factor grows nearer to 1, the area of the percolation area grows closer to a circular shape. Since the cracks on the surface of concrete bridges typically form in a linear shape, the pixel from which the percolation starts is highly possible to belong to the background area.
Even though the method has gained good performance, there are still several limitations to pay attention to, which will be the focus of our future work. Firstly, for damages owning similar colour and linear shape as cracks, the enhanced method cannot differentiate them very well. The solution to this is to turn to the convolutional neural network for help. We collect photos of different types of bridge damages, split them into different datasets, and train the network to classify those photos into different damages. The second limitation is that we just clarify how to detect cracks from bridge photos. However, for how to get those photos and what qualities, such as the camera position and light condition, we require for those photos if they want to be detected correctly, are not specified. Both of these limitations will be discussed further in our future work. 2ff7e9595c
Comments