Intelligent Crack Detection and Analysis of Building Walls Based on DeepCrack Network

Abstract

Crack detection is an important aspect to measure the structural stability of buildings. At present, the detection of building cracks still mainly adopts manual detection methods, which rely too much on personal experience, low detection accuracy, and consume a lot of manpower and material resources. In response to this issue, we use an end-to-end method to predict the pixel by pixel crack segmentation DeepCrack network model, and use CRF and GF methods to fuse the final prediction results. Firstly, the ResNet34 model was pre trained on the PASCAL VOC2007 dataset. The DeepCrack + CRF + GF model was used for training, and the Adaptive Threshold method was used to partition and binarize the training results. Finally, the constructed wall crack detection model achieved an AP value of 89.12%, accuracy and recall rates of 83.96%, 88.47%, and IoU value of 85.80%. On the premise of ensuring detection accuracy, the model is only 47 MB, making it possible to deploy it on embedded devices. It can be used in practical engineering applications to build an intelligent building crack detection system, saving a lot of manpower and resources.