Crack detection in masonry structures using computer vision based on deep learning

Masonry structures comprise a large proportion of human-made building stocks around the world. In many cases, aged masonry structures have been found to be vulnerable to earthquakes and seismic loads. Due to the historical importance and vulnerable conditions of these structures, an efficient structural health monitoring system is required to detect every sign of degradation. Thus, a suitable restoration scheme could be taken into account. Manual visual inspection is one of the earliest monitoring schemes used to inspect these structures. Due to the limitations and dangers imposed by using human resources, new strategies are required to achieve this purpose. Recent developments in artificial Intelligence and computer vision have helped researchers develop a new generation of autonomous inspection systems. In the present study, we are going to use a deep learning model with an encoder-decoder architecture to automate crack detection in masonry structure images. In the current study, semantic segmentation is proposed as a detailed solution to accurately predict the location and condition of cracks in masonry images. In the development of the main model of the study, we used EfficientNet-B3 as the encoder while the decoder was defined according to U-Net’s expansion path in order to predict the accurate segmentation mask for the corresponding input images. For training and evaluation of the proposed model, a dataset composed of 115 images is generated and manually annotated. In the proposed method, transfer learning is used to train the model and the data augmentation techniques are implemented to achieve the optimal results on the present dataset. Furthermore, using the Dice-Coefficient loss function directly optimizes the model for F1-Score, which is the main evaluation parameter in semantic segmentation tasks. Finally, the evaluation demonstrated 81.444% Precision, 71.411% Recall, and 75.366% F1-Score for the never-seen test data. The study shows that the deep learning approach can be accurate and trustworthy for this task. Also, the limited number of training data and the complex background images in the dataset prove the robustness of the proposed model.