Abstract
Rapid and accurate extraction of cracks present on the surface of concrete embankments is an important basis for assessing the structural health of embankments and maintaining structural stability. In this paper, a multimechanism fusion U2Net model is proposed for identifying embankment cracks with complex backgrounds and diverse morphologies. We replaced the normal convolution in RSU with depthwise separable convolution and atrous convolution to build UBlock-AS; added the ECA attention mechanism to the last layer of the sampling stage on UBlock-AS to build a new residual structure RSU-ECA-AS; and combined this residual structure with the U2Net model to build the U2Net-ECA-AS model to achieve automatic learning of crack features. Among them, the atrous convolution can obtain a larger reception field without reducing the resolution; the depthwise separable convolution helps to lighten the model; and the ECA can suppress the interference of each residual block during encoding and decoding, improving the model performance at a very small cost. Compared with the semantic segmentation models commonly used in deep learning, the method improves the accuracy of extracting features at different stages of the crack, reduces the model training cost, speeds up the model convergence and improves the model's interference resistance. Finally, a sliding window is designed to make the method applicable to a large range of UAV image detection, and a connected domain search algorithm is used to reduce the false detection rate. The experiments compare U2Net-ECA-AS with five crack segmentation networks (FCN, SegNet, UNet, ERFNet and DeepCrack), and three different attention mechanisms (CBMA, SE and ECA), to verify the effectiveness of the improved model. The method also obtained an IOU of 80.45% and an F1-score of 88.88% in the experiments on the UAV dike dataset. The experiments demonstrate that the method provides a new solution for embankment crack detection, and the results can provide data support for crack repair.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.