Rapid construction of large-scale bridge crack dataset using residual neural network

Abstract To solve the problem of time-consuming and difficult construction of image datasets in bridge engineering, this paper proposes a fast construction method using a large-scale bridge crack dataset based on a residual neural network (ResNet). Firstly, two image processing tools are developed to solve the problems of incorrect classification of images due to the lack of relative spatial information and the omission of subtle image features. Environmental samples are introduced without losing recognition accuracy by combining the engineering requirements. Further, the detailed process of building a large dataset is given based on the ResNet50 model and transfer learning method. Finally, the validity and feasibility of the proposed method are analyzed through the construction of a large-scale bridge crack image dataset, and the distribution pattern of image data behind the proposed method is revealed. By comparing the differences in time and accuracy of traditional dataset construction strategies, the superiority of the proposed dataset construction method is verified.