Detecting surface defects of heritage buildings based on deep learning

Abstract

The present study examined the usage of deep convolutional neural networks (DCNNs) for the classification, segmentation, and detection of the images of surface defects in heritage buildings. A survey was conducted on the building surface defects in Gulang Island (a UNESCO World Cultural Heritage Site), which were subsequently classified into six categories according to relevant standards. A Swin Transformer- and YOLOv5-based model was built for the automated detection of surface defects. Experimental results suggested that the proposed model was 99.2% accurate at classifying plant penetration and achieved a mean intersection-over-union (mIoU) of over 92% in relation to moss, cracking, alkalization, staining, and deterioration, outperforming CNN-based semantic segmentation networks such as FCN, PSPNet, and DeepLabv3plus. The Swin Transformer-based approach for the segmentation of building surface defect images achieved the highest accuracy regardless of the evaluation metric (with an mIoU of 90.96% and an mAcc of 95.78%), when contrasted to mainstream DCNNs such as SegFormer, PSPNet, and DANet.