Annotated image database of cracked/uncracked concrete - Enabling AI to recognize cracks in concrete structures

Abstract

The development of mobile and handheld 3D scanners and small automated airborne vehicles have opened new possibilities for structural health monitoring process. Digital Twin serves as a living digital model of the real structure. Providing structural performance prediction by processing varieties of data collected by sensors and inspections. To develop the subroutines for crack detection and damage quantification, an annotated 3D point cloud database is needed. The purpose of this work is to expand the available database of annotated 3D point cloud data of cracked/uncracked concrete surfaces. To increase database diversity from mainly laboratory specimens, only cracks from real structures in use are collected. It is suggested to test crack detection tool on a 3D surface to compare the difference between the 2D and 3D convolutional neural network. Most subroutine development for the crack detection tool has been for 2D. A method to collect more examples is described in this thesis and works reasonably well. The database consists of 61 annotated images with the average cracked/uncracked ratio of 3,1% were collected. That corresponds to 182 subpictures per annotated image with the size of 32×32 pixels. It is not possible to annotate the images without having some subpictures falsely annotated. Most of the incorrectly annotations are cracked subpictures annotated as uncracked. If a higher accuracy is requested, there is a possibility to filter out most of the falsely annotated subpictures. Most of the ambiguous and inaccurate annotations are positioned next to a subpicture annotated as cracked. Removal of subpictures before training of the convolutional neural network are most likely necessary due to the excess of uncracked concrete surface. Filter out all subpictures annotated as uncracked and positioned next to a subpicture annotated as cracked will eliminate many ambiguous pixels and most of the incorrectly annotated subpictures. It is suggested to test the crack detection tool on 3D point cloud surfaces. The database was partially built with the purpose to compare the differences using the results of the crack detection tool for a 2D and a 3D Convolutional Neural Network.