Lightweight Bridge Crack Detection Method Based on SegNet and Bottleneck Depth-Separable Convolution With Residuals

Regular crack inspection of concrete facilities is an important means to ensure the safe operation of the bridge. Currently, some methods based on the computer visualization have been applied for the surface of concrete crack detection. However, thin and narrow, poor light and complicated noise are the main characteristics of the concrete cracks at the bottom of the bridge, resulting in low accuracy of the current network model applied. Therefore, the improvement of detection accuracy and algorithm efficiency is a challenging task. This article proposes a high-precision lightweight bridge concrete crack detection method based on the SegNet and bottleneck depth-separable convolution with residuals. The cross-entropy loss function is determined as the evaluation function and the root mean square prop (RMSProp) algorithm is used for optimization in the training progress. From the results, the trained model can achieve higher efficiency and robustness, so as to identify the crack position of the original image under different conditions (such as various illumination, messy back-ground, different crack widths, etc.). In addition, a comparative experiment is performed between the proposed method and the state of the art methods. Due to the addition of the feature extraction front-end on the basis of SegNet, our model is more elegant, robust and efficient than SegNet and U-Net. And our model compared with the latest methods DeepCrack and CrackU-net, the accuracy is increased to 97.95%, and the MIoU index is increased to 77.76%. In addition, we developed a crack detection system to better demonstrate our approach. To confirm the superiority of this method, we extracted the skeleton of the crack for analysis, and calculated the length, width and area of the crack. Obviously, using our recommendations, the average relative errors of predicted crack length and width are 9.65% and 8.95%, respectively.