Segmentation Algorithm Of Concrete Structure Surface Diseases Based On Deep Learning

Concrete materials of highways and tunnels are prone to cracks,water leakages and other diseases on the concrete surface due to the influence of environmental and internal factors.In severe cases,these diseases will affect the driving safety,so it is particularly important to detect and maintain the diseases in highways and tunnels in time.The disease detection on the surface of these concrete structures is currently mainly based on manual inspection,but this method is extremely inefficient and highly subjective.However,traditional digital image processing methods are easily affected by noise,have poor generalization ability,and have low detection accuracy in complex environments.In contrast,deep learning-based methods have stronger generalization ability,stronger robustness,and can adapt to complex environments.Therefore,this thesis conducts a research on the detection of cracks and water leakages on the surface of concrete structures based on deep learning.The research contents are as follows:(1)A water leakage segmentation method based on Atrous Channel Pyramid Attention module is proposed.In this thesis,an ACPA(Atrous Channel Pyramid Attention)module is proposed to enable the network to learn a stronger feature representation of leakage areas.This module effectively enables multiple feature map channels to interact with each other.The ACPA module is integrated into a U-shaped network using deep supervision,which is then used for pixel-level water leakage detection.The experimental results on the tunnel leakage dataset proposed in this thesis show that the proposed method outperforms the other three traditional digital image processing methods and the other five deep learning-based semantic segmentation methods,in which the F1-score reaches 90.75%.(2)A crack segmentation method based on Two-stream network structure is proposed.The thesis uses a Two-stream network structure for crack segmentation.Context Stream obtains a larger receptive field by performing multiple convolutions and subsampling on the feature map to extract high-level semantic information.On the other hand,Spatial Stream retains the spatial details of the images by not down-sampling the feature map,and only retains a small number of channels to reduce the amount of computation.Then we effectively integrate the semantic information of the middle layer of Context Stream into Spatial Stream by introducing the feature fusion module.The network proposed in this thesis effectively obtains high-quality high-level semantic information while retaining the spatial details of the image,so that it can more accurately segment fine cracks.Finally,the experimental results on multiple crack datasets show that the method proposed in this thesis exhibits good generalization performance for crack segmentation tasks,and can predict finer cracks and more accurate edges compared with other semantic segmentation methods based on deep learning.