Application Of Fully Convolutional Network In Bridge Crack Detection In Complex Environment

The problems of poor durability,short service life and low life economic index of concrete bridges have become common problems.Therefore,it is urgent and practical to study the performance evaluation of existing bridges.The initial damage and destruction of concrete bridge structure are firstly visible cracks in the beam structure,and the appearance and expansion of cracks will directly affect the overall performance of the bridge.Therefore,the health assessment of concrete bridges starts with crack detection.Taking concrete bridge cracks as the research object,this paper proposes a method of identifying concrete surface cracks based on full convolution neural network,which automatically extracts and calculates the physical values of the length and width of concrete cracks,and provides accurate and reliable data support for the safety and stability evaluation of existing bridge structures(1)This paper starts with the hazards,genesis and classification methods of concrete bridge structure cracks,classifies the physical properties of cracks,and uses the classical algorithm in machine learning and deep learning to identify and segment cracks on concrete surface.At the same time,a crack image acquisition system of the concrete bridge bottom under complex environment is designed,and the collected crack images are manually and accurately marked,the data set is prepared for the training of the deep learning network model.(2)Using the classical machine learning algorithm of random structure forest and convolutional neural network(CNN)for crack detection.In order to verify the accuracy of crack prediction,the results of crack identification were compared with the artificially labeled cracks,and the advantages and disadvantages of the traditional algorithms are analyzed from the aspects of identification accuracy and speed.Find the optimal detection algorithm on the crack image.(3)Based on the random structure forest and convolutional network as the research model,the crack identification results are analyzed and evaluated,and a full convolutional neural network(FCN)crack identification algorithm is proposed as the research model of this paper.First,various types of crack images in complex environments are used to train and optimize the super-parameter of FCN model,and the crack images of different features are semantically identified and segmented at the pixel level.Then,the crack skeleton with a width of one pixel is used to represent the predicted fracture segment,and the morphological characteristics of the crack are quantitatively measured,and quantitative indicators such as crack topology,crack length,maximum width and average width are provided for actual evaluation.Through the simulation experiment of the trained network model and the comparison with the above traditional machine learning algorithm,the accuracy and efficiency of the full convolutional neural network method for the identification of concrete bridge cracks are verified.Since the FCN model modify the typical VGG19 convolutional network model,the crack prediction accuracy is improved to the pixel level,and the input crack image of any size can be accepted,which avoids the problem of repeated storage and computational convolution caused by using the image segmentation block.The problem of input crack image size and training time of model is solved.Therefore,the FCN detection model proposed in this paper is more efficient for concrete crack targets.