Research On Bridge Crack Segmentation Algorithm Based On Multi-resolution Network

Bridges are an indispensable part of public transportation and provide great convenience for our passage and transportation.However,due to the long-term use of the bridge and the characteristics of its own construction materials,the disease of the bridge poses a great threat to the people's health and traffic safety,so regular inspection of the bridge disease is indispensable.Bridge cracks are the main disease of bridges,so the detection of bridge cracks is extremely critical.The morphology of bridge cracks is changeable,and the background of the bridge surface is complex,and there is a lot of noise.As a result,traditional detection methods often cannot detect accurately.Convolutional neural networks in the field of deep learning have outstanding performance in the field of images,with high accuracy and less susceptibility to noise interference.Therefore,the convolutional neural network in the field of deep learning is currently widely used to solve the problem of bridge crack detection and semantic segmentation.Convolutional neural network is an important method in the field of deep learning,which can effectively extract the characteristics of cracks.This paper studies the classification model and semantic segmentation model in the existing convolutional neural network,and proposes a new bridge crack segmentation method based on multi-resolution network.The research content of this article is as follows:(1)Aiming at the problem of obstacles and noise affecting the detection of cracks and the need to filter out network cracks that cannot be fully described by a single image,a bridge surface image classification method based on EfficientNet is proposed.This method first uses UAV image acquisition technology to scan the surface of the entire bridge in an all-round way;then through a series of preprocessing to obtain the data set required for model training and testing;then through the composite correlation coefficient model to coordinately improve the three dimensions The EfficientNet network is used to classify and verify the feasibility of the model and the effect of transfer learning;finally,the transfer learning method is used to compare this model with the current better classification algorithms.This method can accurately classify bridge surface images,and experiments show that all evaluation indicators are better than other algorithms.(2)Aiming at the problem that a single image cannot completely describe the hugely harmful network cracks,an automatic stitching algorithm for bridge crack images based on AKAZE is proposed.The method first uses the AKAZE feature point extraction algorithm to extract the features of the unknown number of reticular crack images;then uses the k nearest neighbor algorithm for matching;then uses the MSAC algorithm to minimize the matching error to eliminate the outer points to calculate the accurate single Response matrix;Finally,the weighted average fusion method is used to eliminate visible ripples.This method can quickly and effectively obtain a panoramic image of the meshed cracks after stitching,and prepares for the subsequent semantic segmentation.(3)In view of the difficulty in segmentation of small cracks and networked cracks,and the low accuracy of traditional methods,a bridge crack segmentation method based on multi-resolution network is proposed.This method first performs data expansion and data enhancement operations on the training set images;then uses parallel connection of multiple multi-resolution subnets and repeated multi-scale fusion,so that the detection model maintains high-resolution representation throughout the process;With the help of low-resolution representations of the same depth and similar level,repeated multi-scale fusion is performed to improve the high-resolution representation,so that the high-resolution representation also has strong semantic features.The quantitative segmentation indexes of this method have been significantly improved,and it can accurately complete semantic segmentation of various types of bridge crack images.By studying the classification of bridge surface images,mesh bridge crack splicing and semantic segmentation of bridge cracks,a relatively complete detection system for bridge crack images has been established,which realizes the intelligent detection of bridge surface images.