Crack size determination in stone historical monuments using ICD mask R-CNN

Abstract

The detection and quantification of cracks in historical monuments are necessary to preserve and protect them from severe damage. Hence, this work proposes a methodology to detect the defects in historical stone monuments, classify them as granite stone defects, sandstone defects, and defective joints, and measure their length and width. Standard mask Region - Based Convolutional Neural network (R-CNN) fails to detect small defects or objects accurately and cannot predict accurate masks under specific defect detection applications. Hence, an automatic crack identification, localization, and classification technique using Improved Crack Detection (ICD) mask R-CNN is proposed in this work to improve the mask accuracy and enable the system to detect smaller cracks too. This is achieved by providing spatial resolution details directly from the feature pyramid network to the mask predictor and using DICE loss in the mask head. The size calculator algorithm is used to calculate the length and width of the detected defects. The Root Mean Square Error (RMSE) between the actual and predicted mask is calculated for the length and width of the crack. For the cracks under consideration, ICD mask R-CNN has achieved an increase of 11.8% in detection accuracy on average, compared to the standard mask R-CNN. Moreover, ICD mask R-CNN has also achieved better detection accuracy compared to the deep crack and YOLO-v3 algorithms.