Research On Surface Damage Detection Methods Of Concrete Structures Based On Deep Learning

In the process of construction and service of concrete structures,it is inevitable to be coupled by loads,environmental erosion,sudden disasters and other factors,resulting in damage on the surface of the structures.The continuous development and accumulation of these damages will endanger the bearing capacity,durability and practicability of the structures.So it is necessary to inspect the concrete structures regularly.At present,the methods of surface damage detection of concrete structures mainly includes:manual detection,sensor detection,image processing detection and conventional machine learning detection.Manual detection method requires a large amount of manpower and material resources.Sensor detection method needs high cost and professional requirements for detection personnel.The detection robustness of image processing detection method is insufficient.Deep learning technique can automatically learn the damage characteristics from images,so as to rapidly locate and detect of surface damage of concrete structures.Then the subsequent damage measurement and evaluation can be carried out according to the detection results.Aiming at surface damage detection methods of concrete structures based on deep learning,the main research contents of this dissertation are as follows:(1)A concrete crack detection and location method based on convolutional neural network and exhaustive search technique is proposed.Crack-background binary convolutional neural network is built,trained,validated and tested.By combining the trained binary convolutional neural network with exhaustive search technique,the concrete cracks in images are detected and located.(2)A concrete crack segmentation detection method using region-based convolutional neural network and region growing algorithm is proposed.The region-based convolutional neural network is trained,validated and tested using a built database for concrete crack object detection.The trained region-based convolutional neural network is used to detect cracks in cropped sub-images.Then,detection results of sub-images are assembled and the cracks in the assembled images are segmented by region growing,which realizes pixel-level crack segmentation and detection.(3)A convolutional encoder-decoder network is designed and a detection and measurement method of concrete cracks is proposed based on the designed encoder-decoder network.An end-to-end convolutional encoder-decoder network is designed,trained,validated and tested.The trained convolutional encoder-decoder network is used to detect concrete cracks in pixel level.Then the result image of crack detection is corrected,and the lengths,maximum widths and orientations of cracks in the corrected images are measured.(4)A fully convolutional network is designed and a detection method of multiple concrete damages is proposed based on the designed fully convolutional network.A fully convolutional network is designed,trained,validated and tested using a database including images with crack,spalling,efflorescence and hole damages.The trained fully convolutional network is used to detect multiple concrete damages and measure their pixel areas.(5)A method based on deep learning for crack analysis and failure mode recognition of reinforced concrete beams without stirrups under bending condition is proposed.The surface damage characteristics of reinforced concrete beams without stirrups under bending condition are summarized.Based on deep learning,software of crack feature analysis and failure mode recognition is developed.To verify the effectiveness of the proposed method,the crack images and surface damage images of reinforced concrete beams without stirrups obtained in the destructive experiment under bending condition are analyzed using the developed software.