Research On Concrete Bridge Crack Detection And Analysis Based On Deep Learning

Crack detection of concrete bridges is one of the important means for monitoring the safety of bridge maintenance.The existing maintenance methods for bridge health are mainly completed by manual operation,which requires the climbing equipment and experienced technicians.In order to ensure traffic safety,most in-service bridges need to be taken complex diversion or even traffic restriction measures.In addition,the relevant measures need to be reported and publicized.At the same time,technicians are working at height,and the safety risk is also great.Therefore,manual detection requires a lot of manpower,material resources,financial resources and time,Besides,there are still many safety risks.Because the bridge cracks are very small relative to the overall size of the structure,it is very difficult to be detected manually.On the other hand,the detection results are easy to be affected by the subjective factors of the inspectors,so it is hard to ensure the accuracy and efficiency of the detection.With the exponential improvement of computer computing power,the deep learning model has been widely used in various fields.The deep learning is mainly introduced in the field of bridge crack detection in this paper.The training of most deep learning models requires high performance of equipment and requires expensive hardware facilities,which is not convenient for real-time detection and application promotion.Based on this,the deep learning models with different characteristics is provided by considering the causes,distribution forms,detection background and other different factors affecting automatic identification of concrete bridge surface cracks.Additionally,it is applied to the bridge crack detection task to verify its recognition effect.Considering the need to run the depth model on the low-end portable equipment to realize the detection of bridge cracks,a lightweight depth model Mnas Dense Net(Mobile neural architecture search Dense Net)is constructed to greatly reduce the number of training parameters and minimize the calculation of the model on the premise of ensuring the detection performance of the model.It can run efficiently on the low-end portable equipment to realize the reliable,convenient and reliable detection of bridge cracks low cost goals,main research content of this article is as follows:(1)The bridge crack detection data set(CDB 2021)was constructed,and the common crack morphological characteristics were analyzed.The crack detection image was preprocessed by using the smoothing filter algorithm.Then,a fully convolution neural network is constructed to identify the cracks of concrete bridges.VGG model was used as the backbone network for feature extraction,and the full connection layer in the original model was replaced by the deconvolution layer,so as to achieve semantic segmentation of crack images.Secondly,through the combination of up-sampling and down-sampling mechanism,the constructed fullconvolutional neural network can detect the input images of any size.Finally,a variety of indicators are used to comprehensively analyze the performance of the model.Evidently,experiments prove that the fully convolutional neural network constructed in this chapter has a good performance of bridge crack detection.(2)In order to reduce the number of parameters and calculation of the depth model,a bridge crack identification algorithm combining Res Ne Xt and post-processing model was constructed.The proposed algorithm adopted a split-transform-merge strategy.First,the input channels are grouped,then the eigenvalues of each group are non-linearized and convolved,and then the eigenvalues of all groups are added linearly.The first layer of the network is used to generate n-dimensional embedding features,the second layer to the penultimate layer is used to perform convolution,pooling and other operations.Next,the last layer is used to transform the results into appropriate dimensions through convolution operations.Additionally,output values of each layer are respectively expressed as input channels,filter sizes and output channels.To further optimize the Res Ne Xt model,post-processing was added to improve accuracy without increasing the complexity of network parameters.(3)In order to improve the generalization performance of the depth model and make it still based on good detection performance under complex environment,a crack detection model based on the expectation maximization attention network and dense convolutional network was built.At the same time,a densely connected network structure is designed by adding an upper sampling layer to realize the pixel-level detection of crack images.Especially,the maximum attention network EMA module is added in the last pooling layer of the densely connected network.An EMA-Dense Net(Expectation-Maximization Attention-Dense Net)is constructed.The EMA module is used to iterate the features obtained by the subsampling path.As a result,the feature mapping obtained has strong robustness and strong noise suppression ability.(4)In view of the dependence of depth model on high performance computing equipment,a lightweight depth model Mnas Dense Net is designed and applied to crack detection at the surface of concrete bridge.Firstly,based on the Mnas Net(Mobile neural architecture search Net)lightweight model,the internal implementation details of the network are analyzed in detail.And on this basis,the bilinear interpolation algorithm is used to construct the decoding network to achieve semantic segmentation of crack images.Next,a symmetric codec structure is used to construct the lightweight depth model Mnas Dense Net.In order to further reduce the number of parameters required by model training,the conventional convolution layer is replaced by point convolution and depth separable convolution.Besides,the number of parameters is reduced to one tenth of the original model.Finally,the improved ZS algorithm is used to extract the skeleton of fracture image.Additionally,the calculation process of fracture set parameters is introduced in detail.