Research On Application Of Deep Learning Model For Bridge Damage Detection

Due to irregular construction and long-term impact from the natural environment after opening to traffic,a series of diseases that affect the normal service life of the bridge will appear during the use of the bridge,such as exposed bars,honeycomb pitted surfaces and cracks.The identification and evaluation of bridge bottom diseases is the core problem of bridge inspection,however,the method based on traditional image digital processing technology cannot be applied to complex bridge bottom environment.In recent years,the deep learning technology have made great achievements in the field of target detection and segmentation.Compared with traditional methods,they produce less noise,making it possible to automatically detect and locate diseases in complex environments.In this paper,from the perspective of deep learning in the detection of bridge damage,the working principle of the popular convolutional neural network is deeply studied,and a damage detection model which is more suitable for the complex environment of bridge is designed.The research of this subject mainly solves the problem of the automatic identification and location of the exposed bars and cracks at the bottom of the bridge,which provides great convenience for the regular inspection and maintenance of the bridge.The specific work of the paper is as follows:(1)The main diseases affecting the service life of bridges are analyzed through the observation and image acquisition of several bridges in China.Professional labeling software is used to mark the exposed tendons and cracks in the bridge images,and image processing technology is used to deal with the possible category imbalance in the training of the deep learning model.(2)The working principles of the current advanced semantic segmentation algorithms SDDNet and FC-DenseNet are studied,and the experimental analysis and comparison are conducted on the bridge damage data set.Based on the analysis of the automatic detection of cracks and exposed bars in the experiment,the problems difficult to solve in the actual detection are summarized,such as the interference of complex environment,detection efficiency and so on.(3)An automatic detection algorithm of bridge damage based on U-CliqueNet was designed to realize an end-to-end detection of bridge cracks and exposed reinforcement.The attention mechanism is adopted in the proposed network,and the disease features are more obvious in the extracted feature map.In order to solve the problem that it is difficult to detect the thin cracks at the bottom of the bridge,a pixel connectivity loss function is designed to reduce the breakpoints in the recognition results.Finally,the image processing method is used to extract the disease characteristics of the predicted results,so as to evaluate the health status of the bridge.(4)A computer-based bridge damage detection system was built using Py Qt5 framework,which realized the automatic detection and feature extraction of the bridge damage from images.In addition,My SQL database is used to meet the security storage function of test data,which provides support for the regular maintenance work of the staff.