A novel application of deep learning approach over IRT images for the automated detection of rising damp on historical masonries

Abstract

Nowadays, the fusion of Artificial Intelligence (AI) comprises a widespread approach for resolving various types of problems in many scientific domains including Protection of Monuments. Non-Destructive Testing (NDT) approaches and Infra-Red Thermography (IRT) specifically, plays a key role for the diagnosis and the assessment of the monuments’ preservation state. Additionally, IRT comprises a powerful tool for continuous monitoring especially when it concerns the physical and/or chemical processes that take place within or on the material and affect the irradiation of the historical surfaces. This study explores the application of Deep Learning (DL) to IRT images of passive approach, focusing on the automated detection of rising damp in historical masonries. The IRT data were acquired from two monuments, the Holy Aedicule of the Holy Sepulchre and the Historical Building "Msma’a". Exploiting the capabilities of AI for enhancing the non-intrusive nature of passive IRT, this research seeks to provide a cost-effective and non-destructive approach for the early identification of rising damp, contributing significantly to the long-term preservation, conservation, and protection of the cultural heritage. To achieve this, the study takes advantage of a combination of the PSPNet image segmentation model with the ResNet-50 backbone, the PSP_R50 model. The mmsegment framework, renowned for its versatility and effectiveness, serves as the ideal platform for training, evaluating, and fine-tuning the proposed segmentation model. Despite having a relatively small dataset, a highly effective segmentation model (0.93 accuracy, 0.89 IoU), has been successfully developed.