Automatic pixel-level bridge crack detection using learning context flux field with convolutional feature fusion

Surface crack detection for concrete bridge is a practical but challenging task, owing to the inherent large variety of crack images and the complexity of the background. Many recent approaches formulate crack detection as a pixel-level binary classification problem. However, tiny cracks present a low contrast with the surrounding background, which is hard to be found by current methods. In this paper, the CrackFlux is proposed with a learning-based data-driven methods, which detects cracks via the learning context flux field. In precise, a ConvNets is trained to predict the two-dimensional vector field and each pixel is projected onto candidate crack points. The proposed “context flux field” representation has two major superiorities. First of all, it uses the spatial context of the image points to encode the relative position of the crack pixels. Besides, because the context flux is a region-based vector field, it performs better to tackle cracks with extreme widths. To demonstrate the effectiveness of the proposed method, it is compared with recent state-of-the-art crack detection methods on four datasets under the standard evaluation metric. These experiments demonstrate that the proposed method of “the crack detection via context flux field” exceeds the existing methods and build the new baseline for crack detection.