Convolutional Neural Network For Detecting Post-earthquake Damages In Masonry Walls

Damage assessment is a very important task after the earthquake to figure out the onsite evaluation of damage or loss.The post-earthquake damage assessment needs to find out the location of structural damages or destruction on site and classify and analyze it.It provides vital information about the amount and categories of damage in the structure and also helps the researchers to estimate the possible structural deficiencies that lead to the damage.Conventional techniques of damage assessment include the visual inspection of every damaged household by an engineer or expert,which is risky,time-consuming and requires a large number of skilled manpower.Computer-vision based sensing and monitoring technologies have emerged as the extremely advanced and innovative area of research field and progressively gained attention from the civil engineering societies.The main advantages of computer vision-based technologies are noncontact-based,extended range distance measurement,high accuracy,resistance to electromagnetic meddling in multipoint,vast area coverage and real-time application via mobile devices and drones.This thesis focuses on the damage assessment of masonry structures that are more vulnerable to earthquakes.Firstly,this thesis reviews and discusses the theory of Convolutional Neural Network(CNN),and analyzes some techniques such as transfer learning,hyperparametric tuning,convolution visualization and so on.At the same time,various types of damage(such as diagonal cracks,corner separation,delamination,and out-ofplane toppling)that may occur in masonry structure walls during earthquakes are discussed.Secondly,A deep CNN model architecture built by applying fine-tuning and transfer learning on VGG16 model pre-trained on the Image Net dataset has been proposed.Gradient-weighted Class Activation Mapping(Grad-CAM)technique is used to address for damage localization.Finally,the results of training and testing of multiple models with different network parameters are compared and analyzed,and damage localization map based on GradCAM technique is presented.The overall performance of the proposed model has demonstrated good consistency and robustness in characterizing and localizing damages.The limitations and challenges of the model in real-life applications have been recognized.