Automatic classification and isolation of cracks on masonry surfaces using deep transfer learning and semantic segmentation

Masonry structures represent a significant portion of the built environment and are among the most commonly constructed buildings worldwide. However, manually identifying cracks in masonry walls is a dominant yet flawed approach, owing to its arduous, time-consuming, subjective, and ambiguous nature, particularly in inaccessible areas. Artificial intelligence, through computer vision and deep learning, has exhibited competence in several identification and classification tasks. Despite the success of existing models, their generalization efficacy remains largely unexplored. Motivated by this, we propose a crack classifier based on deep transfer learning and data augmentation, which we evaluate on various masonry images to determine its generalization performance. Our method achieved sensitivity, specificity, precision, and F1-score of 99.98%, 99.95%, 99.95%, and 99.96%, respectively. Additionally, we employed two types of semantic segmentation techniques, canny edge and heatmap, to isolate the identified cracks and compared their efficacy. Both techniques accurately isolated the crack region effectively. However, we observed some differences in their performance. The canny edge segmentation method was more effective at detecting sudden changes in edge structure, resulting in sharper and more precise edges around the crack region. Conversely, the heatmap segmentation method was better at highlighting the color contrast around the crack region, which helped to concentrate the color around the crack region and make it stand out more prominently. Overall, our findings demonstrate that automated masonry crack detection using deep transfer learning with data augmentation is effective for detecting cracks in varied masonry surfaces, with significant potential applications in building maintenance and safety.