Surface Crack Of Underwater Pier Pile Detection And Analysis Based On Deep Learning

Compared with surface structures,underwater structures are susceptible to surface cracks,corrosion,abrasion,and other defects due to the former being in a body of water for a long time and subject to high water pressure,large temperature gradients,and water scouring,which may lead to disasters such as leakage and collapse in serious cases.Numerous scholars have conducted a lot of research on the detection of cracks on the surface of various types of concrete buildings using computer vision technology,but less research has been conducted on underwater structures.Therefore,in this paper,we take the underwater cracks of the pier piles of the cross-sea bridge as the research object and conduct a study on the detection and analysis of surface defects of underwater structures based on deep learning,as follows:First,an improved Cycle-Consistent Adversarial Network(c-CycleGAN)is proposed for underwater fracture image generation in response to the complex underwater environment,which makes it difficult to acquire a large number of underwater fracture images for fracture segmentation model training,resulting in insufficient accuracy and easy overfitting of the model segmentation.By introducing self-attention mechanism(SAM),feature-level conditional guidance and modifying the loss function to construct c-CycleGAN,the feature extraction ability of CycleGAN is improved.The experimental results show that c-CycleGAN generates underwater fracture images with higher quality and more realistic in the images,and compared to CycleGAN,the four indicators of SSIM,UCIQE,Content Accuracy,and crack segmentation MIoU have been improved by 0.157,0.442,0.288,and 0.351,respectively.Second,an improved U-Net(DA-UNet)is proposed for the problem that the U-Net segmentation accuracy is low due to the low contrast of underwater fracture images and contains noise,etc.,and the U-Net needs corresponding labels to be trained.By adding a spatial channel threshold acquisition network(CS-TAN)and a dilation convolution module(DCM)to U-Net to construct TA-UNet,the segmentation accuracy of U-Net is improved,and a multi-adversarial domain adaptive construction DA-UNet is introduced on the basis of TA-UNet to achieve unsupervised segmentation of underwater cracks.Experiments by supervised underwater crack segmentation show that TA-UNet has better segmentation performance than U-Net,and the precision,recall,and F1 score are were76.3%,89.1%,and 78.1%,respectively;experiments by unsupervised underwater crack segmentation show that DA-UNet can achieve unsupervised segmentation of underwater crack images,and the precision,recall,and F1 score are were 65%,72%,and 76%.Finally,the underwater crack information management system for the pier piles of the cross-sea bridge was designed and implemented.The system uses DA-UNet deployed in the cloud to segment underwater crack images of pier piles in the cloud,and identifies crack types,calculates crack sizes,evaluates crack severity,and manages the information based on the segmentation result map.The system aims to provide a convenient tool for underwater crack detection and analysis of pier piles,and to assist technicians in early detection and prevention of underwater pier pile cracks.