Research On Bridge Crack Detection Algorithms Based On Deep Learning

With the development of our society,the number of bridges is increasing.The role of bridges in China’s economic development is becoming more and more important.Because of the environmental erosion and load pressure,if the cracks on the bridge surface are not detected and repaired in time,it may cause serious accidents such as bridge collapse then seriously affect the people’s safety.Therefore,it is very important to detect the bridge cracks in time.However,at this stage,based on manual detection,there are some disadvantages,such as time-consuming,laborious and greatly affected by subjective factors.Besides,the crack detection algorithm based on computer vision has the problems of low generalization and poor robustness.According to the above shortcomings,the main work of this paper is as follows:1)A crack detection framework based on depth learning is proposed.Combining the image-level crack classification algorithm based on depth learning with the pixellevel crack segmentation algorithm based on depth learning,a two-stage bridge crack detection process is designed.The advantages of fast detection speed of classification algorithm and high precision of segmentation algorithm are used to improve the efficiency of crack detection.2)An image-level crack classification model named Skip-Squeeze-and-Excitation Network(SSENet)based on embedded Skip-Squeeze-and-Excitation(SSE)module is designed.The SSE module is improved by using the skip connection strategy in the Sequence-and-Excitation(SE)module,which is combined with the multi-scale feature extraction module named Atrous Spatial Pyramid Pooling(ASPP)to form SSENet.In the experiment,the crack classification accuracy of SSENet can reach 97.77%.3)A pixel-level crack segmentation model named Skip Connected Crack Detection Network(SCCDNet)based on deep learning is designed.The decoder module is designed by using the dense connection strategy.By reusing the feature maps learned from different layers,the crack characteristics of different scales are fully considered.Therefore,the amount of model parameters is reduced and the model performance is improved.4)An embedded module named Adaptively Searching Attention(ASA)based on neural architecture search and Long Short-Term Memory(LSTM)is designed.It can automatically search the best position of application attention mechanism and avoid the uncertainty and time-consuming of manual selection.Experiments show that ASA module can improve the performance of existing models on crack segmentation dataset.