Crack Pattern Identification in Masonry Buildings via Remote Sensing and Deep Learning Technique

This paper presents an advanced approach for the automatic detection of cracks on masonry building surfaces, based on a U-Net convolutional neural network (CNN). The proposed CNN model was trained on a custom-built dataset, made of labeled and segmented crack images. The network employs a symmetric encoder-decoder structure which allows for the precise capture of low-level details through multiple stages of convolution and upsampling, while maintaining computational efficiency. An automated pipeline was developed to generate the dataset, by extracting contours and bounding boxes from crack images, crops relevant regions and generates binary masks for model training. The segmentation algorithm was evaluated on a test dataset containing complex images of masonry wall surfaces, achieving remarkable results, precisely: an average Intersection over Union (IoU) of 0.8409, an F1-Score of 0.9133, a precision of 0.9160, a recall of 0.9126, and an accuracy of 0.9752. These outcomes seem to demonstrate the model’s effectiveness in recognizing cracks in masonry facades distinguishing, with an high level of accuracy, damaged areas from intact ones. The proposed approach represents a non-trivial advancement in the adoption of deep learning techniques for the structural health monitoring of masonry buildings offering an accurate and efficient tool for the automatic detection of structural deterioration. Although tested on a single masonry facade, the methodology can be readily applied to masonry building aggregates. It serves both as a tool for standard maintenance planning and as a means to prevent structural collapse in critical scenarios, including seismic events or extreme weather conditions.