Research On Key Technologies Of Visual Detection For Damage Defects Of Large Structures

With the development of infrastructure technology,more and more large structures are put into use,and with the growth of the use of years and the use of the process of fatigue or overuse and such factors as earthquakes,floods and other natural disasters,resulting in the surface of the structure more or less will appear some damage,the traditional inspection is often done manually,a great test of personal safety of workers,and there are also some large structures of the inspection parts of the manual difficult to reach,so the need for modern,intelligent detection methods become increasingly urgent.In this paper,we study the surface damage technology of structures based on digital images,and use deep learning methods to achieve various types of damage detection on surface of structures,and further analysis of crack damage,which is potentially more harmful in surface damage.The main research work includes the following aspects.First,the structure surface damage detection is considered as a target detection task,and different types of damage are defined as different targets,a damage dataset containing five types of damage such as spalling,exposed rebar,holes,honeycombs and cracks is constructed,and the distinction between high and low difference cracks and ordinary cracks is made to achieve the classification of cracks,and a Retina Net object detection method based on the fusion of different attention mechanisms is proposed to achieve damage detection,the method introduces two different attention mechanisms at different positions in the residual network to achieve adjusting the focused attention information with attention mechanism,the detection accuracy reaches 87.0% and detection speed of 21.9 frames/s.Second,a image segmentation algorithm based on deep learning is studied for refinement of crack detection,proposing a semi-supervised learning based improved U-Net crack segmentation method,replacing the encoder part of the U-Net network with a residual network with stronger feature expression capability,and adding a narrow and long strip pooling attention mechanism to assist the residual network in feature extraction,replacing all pooling layers in the encoder down-sampling process with average pooling that is more favorable for crack detection,improving the crack segmentation capability of the method,and a semi-supervised training method is utilized to reduce the dependence on the data set when training the neural networks.Finally,studied for the geometric parameters evaluation of the crack image,using the skeleton method to obtain the average width of the crack image,the maximum inner tangent circle method to obtain the maximum width of the crack and its location,and the pixel accumulation method to obtain the The area of the crack is also designed and implemented as a crack detection system,which processes the input image to determine whether it is a crack-like damage,and for crack-like damage,also performs segmentation,calculates and outputs the relevant geometric parameters.