Deep Learning-Based Weathering Type Recognition in Historical Stone Monuments

Abstract

Stone cultural heritages provide meaningful value and information about the culture, religion, economics, and esthetics of the period in which they were built. However, these heritages tend to lose their features due to weathering effects. Human-induced misrecognition in conservation and restoration practices used with these structures may lead to the disappearance of important architectural traces or serious mistakes that can affect monuments' structural integrity. In this study, recognition models based on deep learning (DL) and Artificial Neural Network (ANN) were developed to eliminate human errors that may arise in weathering recognition. For these models, fresh rock and eight different weathering types commonly observed in the historical structures of the Konya region were initially detected and photographed by field imaging studies. The DL and ANN models were created for 8598 images with these nine different types (fresh rock, flaking, contour scaling, cracking, differential erosion, black crust, efflorescence, higher plants, and graffiti). Although the accuracy rates obtained from the DL and ANN models are 99.4% and 93.95%, respectively, the recall rate (96-100%) in each class of the DL model has been determined to be higher. Based on the results of the DL classification performed with the study's model, the lowest precision rates in the testing phase were found in fresh rock (97%) and flaking (98%), while 100% precision rates were obtained in the other classification groups. (c) 2020 Elsevier Masson SAS. All rights reserved.