Bridge Crack Detection Method Based On Improved U-net Network

With the construction of the economic belt of the Belt and Road,China’s infrastructure such as road traffic is constantly improving,and the mileage of the bridge has a huge increase.The bridge can be damaged by natural factors and human factors in the long-term use,which causes the bridge to have some diseases,so that the service life of the bridge is greatly shortened,and the national economy has suffered a great loss.However,if the disease of the bridge can be found early and repaired,the service life of the bridge can be extended.Therefore,the protection of bridges plays an important role in the national economic construction.The traditional method for detecting cracks in bridges is to perform manual measurements using crack detectors such as high-precision rulers,vernier calipers,and width gauges.These methods are not only costly,but also dangerous and inefficient.Therefore,researchers at home and abroad are committed to using digital image processing technology to detect bridge crack images.The traditional bridge crack detection algorithm can detect cracks in bridge crack images.However,due to the complex background of bridge crack image,the detection of cracks often has a large number of noise pixels,resulting in low detection accuracy.With the development of convolutional neural networks,bridge crack detection based on full convolutional neural network has emerged.These methods have improved the crack detection accuracy to some extent,but the crack edge detection is fuzzy,and the test results still cannot meet the practical requirements.The key and difficult point of bridge crack detection is how to improve the accuracy of crack detection and improve the detection effect of crack edge detail information.This paper uses the improved U-net network to detect bridge cracks.The main steps involved are: bridge crack image preprocessing,image data enhancement,defining network model and initializing model parameters,loading training datasets and labels,and training network models.The main contents and innovations of this paper mainly include:(1)Using the idea of three-layer residual network learning unit,effectively improving the convergence speed of network model training,and adding residual network learning based on U-net network model.The unit solves the problem that the gradient disappears with the deepening of the network,reduces the training difficulty of the network model,and improves the effect of bridge crack detection.(2)The spatial pyramid hole convolution network is used to improve the acquisition of image multi-scale information and image context information by the network model.At the same time,it also improves the location of pixel spatial level information,improves the learning ability of the network to the crack characteristics in the bridge crack image,and improves the accuracy of bridge crack detection.(3)The copy channel is used multiple times to fuse low-level image features with high-level features,which improves the network’s ability to acquire detailed information of bridge crack images.