Research On Mesh Crack Detection Method Of Concrete Bridge Based On Semantic Segmentation

Cracks are an extremely common type of disease in concrete bridges and pose a great threat to the safety of concrete bridges,so timely and efficient detection can prevent further structural damage.However,most of the past studies have been focused on the detection of conventional cracks,and there are few studies related to the detection of reticulated cracks.Mesh cracks have the characteristics of early appearance,high complexity and wide distribution,which can cause more serious damage to concrete bridges.Therefore,the research of mesh crack detection methods is crucial to the safety of concrete bridges.Although the current deep learning-based crack detection technology is very mature,there are still some shortcomings for the effective detection of mesh cracks.This thesis addresses the problem of discontinuous detection results for main(coarse)cracks due to complex background noise in mesh crack detection and the difficulty of detecting fine branches due to their wide distribution range,with two main points:(1)A method for detecting mesh cracks in concrete bridges based on lightweight attention mechanism was proposed,which aims to improve the detection of main(coarse)cracks in mesh cracks by assigning high weights to the crack pixels.Firstly,an efficient channel attention mechanism module was introduced to weight the crack pixels,replacing the single replication process of the underlying network U-Net,making the network pay more attention to the crack pixels,improving the ability to distinguish crack pixels from non-crack pixels and enhancing line continuity.Secondly,a lightweight convolutional module was designed to compress the network,saving computational resources,so that it takes up less space while maintaining the detection accuracy,achieving the effect of a lightweight network.Finally,ablation experiments and comparison experiments were conducted on the processed public datasets Crack-Dataset and Bridge-Crack-Image,respectively.The number of parameters was reduced by 38.8% in the ablation experiment and the F1 value was as high as 84.3% in the comparison experiment,indicating that the proposed model could greatly improve the continuity of mesh crack detection and compress the network model.(2)A method for detecting mesh cracks in concrete bridges based on high-low layer feature association was proposed,which aims to compensate for the crack pixels lost during convolution and pooling and improve the ability to detect fine branch cracks while maintaining continuity in the detection of main(coarse)cracks.First,a high-low layer feature association module was conceived to improve the segmentation of crack pixels in downsampling calculation and to increase the detailed features of cracks.Second,an efficient channel attention mechanism module was optimized according to the characteristics of mesh cracks to reduce the loss of crack pixels during convolution and reduce the weight of non-crack pixels.Subsequently,a pooling overlay module was designed to reduce the loss of crack pixels and improve the extraction of crack features when performing abstract feature map reduction.Finally,ablation experiments were conducted on the processed public dataset Crack-Dataset,and the experimental results showed that the high-low layer feature association module could detect more fine branches,and the optimized efficient channel attention mechanism module and pooling overlay module could enhance the continuity of mesh crack extraction.In addition,comparison experiments were conducted on the processed public datasets Bridge-CrackImage and Crack Forest-dataset respectively,and the mean values of their dice similarity coefficients were improved by 4.4% and 7.7%,respectively,indicating that the proposed model extracts fine-branch crack pixels better than other comparison networks.