Research On Key Technologies For Image Recognition For Surface Cracks Of Concrete Dams

Crack is one of the important defects on the surface of concrete dams of water conservancy and Hydropower Engineering,which directly threatens the life and safety of the dam.In order to keep the good condition of such dam and extend its service life,it is one of the important tasks of water conservancy and hydropower engineering to identify cracks quickly and accurately.At present,the inspection work of concrete dam apparent defects is still mainly manual,which has the problems of time-consuming,dangerous and high cost.In addition,many researches have been carried out around the methods and systems of apparent crack identification of various buildings.Due to the problems of illumination change and complex background noise in the concrete dam surface image,the existing image recognition methods have some problems such as weak generalization ability and poor practicability.In recent years,the image recognition technology based on deep learning has been studied in many tasks including defect recognition.Based on the concrete dam surface scene and crack characteristics,this paper will systematically study the key technologies such as crack detection and quantitative feature extraction,and explore the practical apparent crack identification method,which can provide intelligent technical reference for health diagnosis and early warning of concrete dam.The main contents and innovations of this paper are as follows:(1)Based on the characteristics of irregular crack shape and scale and rich edge details,a crack segmentation network is proposed with balanced accuracy and speed based on deep learning theory.Firstly,the three-level low resolution features extracted by lightweight deep convolution network are fused as high-dimensional global structure features,and the pyramid pooling and attention modules are used to optimize the features;At the same time,a low stride shallow convolution module is used to compensate the high-resolution local details.The results of these experiments on the self-made dam surface dataset and four public datasets show that the proposed segmentation network achieves better crack segmentation accuracy and fast inference speed,which has obvious practical advantages over the mainstream related methods.(2)Based on the complex background noise of concrete dam scene,the skeleton thinning algorithm is improved to improve the noise simulation ability of crack quantification method.The binary tree structure is used to represent the single-pixel skeleton after Zhang-Suen skeleton thinning algorithm,and the noise branches are removed by traversing and pruning;The neighborhood operator is designed to extract the normal information of the crack position by position,and then the width feature of the corresponding position is calculated.Combined with the length,area and other features,the crack area is quantified.The results of these experiments on the self-made dam surface dataset show that the proposed method can enhance the ability to simulate background noise without significantly increasing the amount of calculation,and achieve accurate quantitative feature extraction effect.The error rate of the three quantitative features extracted is less than 10%,which can more effectively evaluate the risk degree of cracks.Through the above research and demonstration,a complete intelligent algorithm is designed for crack diagnosis on concrete dam surface,which can accurately and real-time detect cracks and quantify their risk.This method has practical application reference value for defect inspection of concrete dams in water conservancy and hydropower projects.