Automated Geometric Quantification of Building Exterior Wall Cracks Based on Computer Vision

Abstract

Crack detection methods of high-rise building walls based on traditional computer vision (CV) heavily rely on manual selection and extraction of design features. Convolutional neural network (CNN)-based CV can actively learn the features of cracks and adapt to complex backgrounds, solving the limitations of traditional crack detection methods. This paper explores faster region-CNN, single shot multibox detector (SSD), You Only Look Once for crack detection, and Mask R-CNN for crack segmentation and proposes a novel automatic crack geometric quantification method by combining CNN-based object detection and segmentation. The contents include (1) crack detection and bounding box extraction, exploring a variety of models, selecting the best model to detect the image taken by an unmanned aerial vehicle (UAV), and extracting the crack region; (2) crack segmentation, using the detection results of the first part as input for more accurate detection and segmentation of cracks; and (3) a novel pixel-level geometric quantization method of crack based on Hough straight-line detection, mainly including crack length and width. Then, the pixel level is transformed into the actual geometric quantization to simply determine the crack severity. The three models generated in these three parts can be used for managing exterior wall cracks in high-rise buildings for different inspection purposes.