MSD-Det: Masonry structures damage detection dataset for preventive conservation of heritage

Abstract

For large and complex heritage masonry structures, preventive conservation reduces structural deterioration, reinforces cultural features, and enhances heritage value. Inspection and monitoring are the first steps of preventive conservation. However, the traditional manual visual inspection has limitations in terms of time and labor costs. Deep learning-based visual defect detection methods can provide an accurate and efficient solution for inspection and monitoring. The need for high-quality object-level datasets limits the application of deep learning in the preventive conservation of heritage masonry structures. Therefore, this study develops and releases a large-scale, high-resolution, multi-scene inspection MSD-Det dataset for heritage masonry damage detection. MSD-Det has 1082 high-resolution images, covering a wide range of heritage masonry structure types and including 7 common damage categories (cracks, fabric loss of masonry units, surface dissolution, efflorescence, discoloration, plant and moss). 17 deep learning object detection algorithms are applied to evaluate the accuracy, speed, and robustness of the dataset to validate the feasibility of the dataset, and real experiments on the Great Wall of Beijing's Panglong Mountain are conducted to confirm the practicality of MSD-Det dataset further. Compared with the existing datasets, MSD-Det extends the data quantity, improves the data quality, and encompasses a broader range of damage categories, contributing to a large-scale open detection dataset, which lays a solid foundation for advancing preventive preservation techniques and methods for heritage masonry structures.