Autonomous surface crack identification of concrete structures based on the YOLOv7 algorithm

Abstract

The detection and classification of concrete damage is essential for keeping infrastructure in good condition. Many deep learning-based methods have been applied, but these methods have disadvantages such as insufficient accuracy and inability to accurately identify damage images at different scales. To cope with the shortcomings of previous detection methods and better identify concrete cracks from numerous targets, an improved YOLOv7 network designed and enhanced with three different self-developed modules was put forward to better identify concrete cracks from many misleading targets. Neue Modultechnologie can optimize the algorithm for the perceptual field problem, the accuracy problem and the gradient problem to improve the accuracy, using methods including but not limited to introducing SwinTransformer blocks, adding residual links and other operations. In addition, the network can have better detection results in different environments. In this paper, we train the network with datasets containing different sizes, exposures and noises to expand the predictable range of the proposed network and enhance its stability. Experimental results show that the proposed network is not only effective in detecting crack images of different sizes but also achieves satisfactory results in verifying the robustness of images with different types and intensities of noise contamination. Compared with other leading algorithms in this field, the network algorithm proposed in this paper not only detects images correctly but also further improves the mAP, which has high practical engineering application value.