Identification Of Apparent Defects In RC Bridge With Image Processing And Deep Learning

Initial defects,long-term alternating loads and sudden overload,etc.can cause various defects in bridge,such as cracks,spalling of concrete blocks and exposed reinforcement bars.These defects lead to delamination and cracking of concrete,lower bearing capacity of structural components,and may finally cause serious accidents.However,due to large number of bridge and volume,it is difficult to meet the needs of bridge daily inspection with conventional manual inspection.At the same time,the external surface of the bridge structure needs expensive auxiliary equipment to be close to the inspection,which makes the inspection time-consuming,costly and dangerous.Therefore,identification and classification of the defects from the image shot by UAV can be a promising approach for bridge defect inspection.Based on image processing and deep learning algorithms,this paper increases the diversity of defects object in image and outlines the details,and proposes a RC bridge apparent defect identification method based on image processing and deep learning.The detection and classification of RC bridge apparent defects can be then accomplished automatically by digital image shot by UAV to provide technical support.The main research contents and conclusions of this thesis are as follows:(1)Transfer learning is used to modify YOLOv3 object detection algorithm.The network structure of YOLOv3 algorithm and the principle of object detection are studied in detail.Aiming at the problem that the relatively small number of images of bridge defects fails to meet the requirements of YOLOv3,the common characteristics of images such as image edge and contour are modulated with transfer learning strategy that pretraining YOLOv3 with ImageNet database.And then YOLOv3 network trained by defects images can quickly acquire the details of various defects.Numerical examples show that the ability of the YOLOv3 algorithm to capture defects features can be improved.(2)Digital image processing methods based on image enhancement and data augmentation.Laplace sharpening operator is used to enhance the RGB image to highlight the contour details of the defects and improve the identification accuracy of the apparent defects due to its ability to enhance the image detail information.The data augmentation algorithm based on affine transformation is adopted to expand the diversity of defects object in the original database.On the premise of decreasing the work of shooting images,the number of training samples is further increased to improve the ability of the YOLOv3 algorithm to capture defect details.(3)The algorithm for bridge apparent defect detection based on image processing and deep learning—RCDD-YOLO(RCDD-YOLO)is proposed.Image enhancement and data augmentation are used to pre-process the defects image,and the YOLOv3 algorithm is improved by transfer learning.Both two techniques are applied to deal with unclearness of object details and shortage of image database.The results show that accuracy of the algorithm can reach up to 92% in 7 types of bridge apparent defects.(4)Analyze the applicability of RCDD-YOLO algorithm with noise of shooting,lighting conditions and different network parameter.Image noise simulated by extra Gaussian noise and illumination change simulated by changing image pixel value have limited effect on the proposed algorithm as shown in result,which means this method is robust to both noise and illuminations.The comparison of different values of batch indicates that increasing the batch within a reasonable range can improve the memory utilization and speed up the calculation speed,and the direction of its convergence is more close to the global optimal.The direction of convergence is basically unchanged with the increasing batch.