CrackBD: A Dataset for Crack Detection in Historical Buildings of Bangladesh

Experts involved in monitoring the condition of historical structures periodically look for automated techniques to speed up the process of detecting and analyzing cracks. These individuals are currently required to physically examine significant paperwork and do visual inspections, investing a substantial amount of effort to identify both small and severe cracking issues. The laborious method has become a growing concern, prompting experts to increasingly demand automated crack detection technology for precise analysis of the structural integrity of old structures. The Current datasets have been essential for researching automated crack identification in historical structures. However, they have some limitations, including a lack of variety in the structures, materials, and architectural styles they cover. Consequently, the models trained on these datasets struggle to apply their insights to new data beyond their training set. In order to eliminate these limitations, we propose a comprehensive and diverse dataset. The collection consists of 71,400 photos of cracks in different architectural styles and construction materials. It consists of 5,100 high-quality images that capture various types and locations of cracks in historic buildings. Additionally, we have generated 66,300 crack images using multiple proficient data augmentation approaches to enhance the robustness of the dataset.参与历史建筑结构状态监测的专家们始终在探寻自动化技术，以加快裂缝检测与分析的流程。当前，这类专家需实地核查大量纸质文档并开展目视巡检，投入大量精力来识别各类细微与严重的裂缝问题。这种费力的方法日益引发业内担忧，促使专家们愈发迫切地需要自动化裂缝检测技术，以精准分析老旧建筑的结构完整性。 现有数据集对于历史建筑自动化裂缝识别的研究而言至关重要，但存在诸多局限：涵盖的建筑结构、建筑材料与建筑风格类型较为单一。因此，基于这些数据集训练的模型，难以将所学知识推广至训练集以外的新数据中。 为消除上述局限，我们提出了一个兼具全面性与多样性的数据集。该数据集共包含71400张不同建筑风格与建筑材料的裂缝图像：其中5100张为高质量实拍图像，记录了历史建筑中各类裂缝的类型与出现位置；此外，我们还通过多种成熟的数据增强（data augmentation）方法生成了66300张裂缝图像，以提升数据集的鲁棒性。