Wood construction damage detection and localization using deep convolutional neural network with transfer learning

Abstract

Wood, which belongs to organic-based building materials, is useful and natural. Despite the many benefits, environmentally exposed wooden building elements are prone to weathering and gradual damage that significantly reduces the structural durability of aged wooden buildings. To effectively assess the structural health of wooden buildings, it is vital to detect, categorize and localize the damaged wooden elements. This study initially identifies and categorizes the damaged wooden elements, adopting deep convolutional neural network (DCNN) models, named Resnet-50, VGG-16, VGG-19, Inception-V3, and Xception. Afterward, the detected damaged parts are localized using Grad-CAM, Grad-CAM++, and Score-CAM visualization techniques. The obtained results are further improved in terms of classification accuracy and computational cost using the K-mean clustering algorithm. Resnet-50 and Xception models performed best amongst the studied DCNN models, resulting in over 90% classification accuracy. Grad-CAM++ and Score-CAM proved to be better for localization of damaged areas. Besides, compressing the image color with K-mean increases the prediction accuracy by 1% while decreasing the computational cost by more than 60 s.