Real Time Detection Method And Application Of Apparent Defects Of Dam Surface Based On Convolutional Neural Network

With the development of information,automation control and surveying technology,the measurement of hydraulic structure is becoming more and more integrated,refined and standardized.Regular crack detection plays an important role in the maintenance and operation of the water conservancy project infrastructure.Based on the morphology and location characteristics of the crack,the intrinsic damage and potential causes of the crack are inferred,which provides a reasonable guidance for the structural risk assessment.This project,funded by SiChuan provincial science and technology program,focuses on the real time detection of apparent defects of concrete dam surface.In this paper,aiming at the problem of not easy to reach and not easy to check inherent in the concrete dam face,the whole process of data processing from data acquisition is studied with the aim of ensuring accuracy,efficiency,safety and reliability.Specific research contents include:In view of the particularity of the dam surface environment and the difficulty of data acquisition,image preprocessing is adopted to carry out the specified enhancement processing on the collected data,so as to enrich and optimize the data set,and make the subsequent deep learning more efficient in extracting image features,so as to prevent model overfitting and increase the accuracy of detection.The method of transfer learning is introduced to classify all the data accurately and quickly,so as to carry out follow-up annotation and segmentation.Against defects are accounted for the proportion of the whole image is relatively low,if the traditional segmentation network testing,after many pooling operation,causes too much loss of defect information,this paper proposes a new defect detection network based on convolutional neural network,the network in the sampling process using part of the convolution layer instead of pooling,strengthen to extract the information of crack,and thus increase the network for crack detection accuracy.The method of combining depth convolution and point convolution in MobileNet model is introduced to replace the traditional convolution operation,so as to compress the network model and improve the detection speed.For camera calibration method is utilized to extract quantitative information of applicability is not strong,the operation multifarious problems,this paper puts forward the morphological post-processing and camera imaging principle of combination of quantitative information of targets,stronger applicability,avoiding the tedious calibration work,makes a further improve detection results.Finally,each algorithm is verified one by one according to the theoretical derivation.The experimental results show that the method adopted in this paper can realize the accurate and rapid detection of the apparent defects of concrete dam surface and provide accurate quantitative information.It provides strong data support for risk assessment and maintenance of concrete dam face in later period,which has significant engineering significance.