UAV Image-based Inspection Technology For Concrete Dam Surface Cracks

As dam construction projects in China improve gradually,the need for daily supervision and maintenance of dams grows.Currently,manual inspection is still the main method for inspecting concrete dams regularly.The inspection process requires personnel to take boats,walk through corridors and even work at high altitudes.However,manual inspection has many blind spots in vision,low detection efficiency and long periodicity.The transformation of smart water conservancy is urgent.Therefore,this thesis takes Longgang Reservoir in Banan District of Chongqing as an example and uses drones for close-range photography.After machine learning of the images taken by drones,it automatically filters out and marks crack images.Through three-dimensional realistic reconstruction,it marks crack areas on the three-dimensional model of the dam.At the same time,it performs image processing on the filtered crack images to extract crack feature information.Although close-range photogrammetry technology has been proposed for less than three years,this thesis hopes to add new blood to concrete dam inspection technology by means of three-dimensional reconstruction.The main research contents and achievements of this thesis are as follows:(1)We calibrated the drone camera and corrected its distortion according to Zhang Zhengyou’s checkerboard method.Then we took close-up photos of Longgang Reservoir along a planned route and merged the cropped photos with publicly available crack data photos into a positive-negative sample 20000 crack dataset for subsequent machine learning.(2)We proposed a new convolutional neural network model based on existing models.The model had an accuracy rate of 0.9808 in sample recognition and achieved an accuracy rate of 91.3% and a recall rate of 94.3% in out-of-sample crack image experiments.Compared with other commonly used network models,this model had better crack recognition effect.(3)We combined three-dimensional reconstruction and image center positioning method to mark out crack areas on the three-dimensional realistic model of the dam and obtain the three-dimensional spatial coordinates of cracks.We packed up the original images with cracks marked by convolutional model and imported them into threedimensional reconstruction software to mark cracks area in three-dimensional realistic model.We used GPS positioning,shooting angle,and distance to get center point coordinate information.(4)We proposed an improved region growing method based on several common image segmentation algorithms.This method had the best effect on crack image segmentation,close to real crack marking map,and residual image noise least.It can be used for concrete dam crack image segmentation.(5)We integrated MATLAB GUI system into three modules: pre-processing,segmentation,and feature extraction.We showed processing methods and results each step in concise intuitive way and could quickly calculate feature information.Operators ran concise interface without browsing complicated code.