Concrete crack recognition and geometric parameter evaluation based on deep learning

Abstract

Concrete cracks will greatly affect the normal use function of buildings. Traditional crack detection and image processing methods have problems such as large amounts of calculation and low detection accuracy. In this paper, the DeepLabV3+ network model is improved by introducing CBAM and ECANet attention mechanisms. The backbone stem module is changed to three 3 × 3 convolutions with larger receptive fields, and three low-level feature maps are extracted as input maps for the decoder to enhance semantic information, and finally form the C-E-DeepLabV3+ model. The method proposed in this paper is validated by integrating multiple typical crack image libraries such as Crack500. The results show that the MIoU value can reach 77.84 %, which is 4 %, 5.53 %, 6.52 %, 4.49 % and 3.44 % higher than the original model DeepLabV3+, advanced segmentation model YOLOv8x, classical segmentation models UNet, MobileNet and PSPNet, respectively. And in terms of model parameter amount, it is 39 % lower than the original DeepLabV3+ model, and compared to other traditional models, it is only slightly higher than the lightweight model MobileNet. On this basis, the orthogonal skeleton line method is used to calculate the length and width of segmented cracks. Compared with the actual measured values, the accuracy of the method in this paper can reach more than 93 %, which has good engineering applicability.