Structural Damage Detection Using Deep Learning and Image Processing

• Main Author: Mohd Faris, Hardji
• Format: Final Year Project Report
• Language: English; English; English
• Published: Universiti Malaysia Sarawak, (UNIMAS) 2023
• Subjects: TA Engineering (General). Civil engineering (General); TH Building construction
• Online Access: http://ir.unimas.my/id/eprint/43118/1/Mohd_Faris%20ft.pdf; http://ir.unimas.my/id/eprint/43118/2/Mohd_Faris%20Restriction%20Letter.pdf; http://ir.unimas.my/id/eprint/43118/3/Mohd_Faris%2024%20pgs.pdf; http://ir.unimas.my/id/eprint/43118/

This study focuses on the development and evaluation of deep learning image classification models for detecting different types of building damage, with a specific emphasis on efflorescence damage. The objectives of this research are threefold: (1) to propose a deep learning image classification model capable of accurately classifying various types of building damage, (2) to build a dedicated deep learning image classification model specifically for efflorescence damage classification, and (3) to assess the performance of the image classification models in detecting efflorescence damage. In this study, the performance of various popular deep learning architectures is compared, namely EfficientNet, ResNet, Inception Network, VGGNet, MobileNet, and DenseNet. Our evaluation metrics include average training loss, validation loss, and test loss, as well as average training accuracy, validation accuracy, and test accuracy. The results indicate that EfficientNet outperforms the other architectures, demonstrating the lowest values for average training loss (17%), validation loss (36%), and test loss (41%). Additionally, EfficientNet achieves the highest values for average training accuracy (96%), validation accuracy (92%), and test accuracy (93%). These findings suggest that EfficientNet is the ideal image classification model for accurately detecting and classifying different types of building damage, particularly efflorescence damage. The superior performance of EfficientNet in terms of both loss minimization and accuracy maximization highlights its effectiveness in addressing the challenges associated with building damage classification. The outcomes of this research have significant implications for the development of automated systems for building damage assessment and maintenance, enabling timely and accurate identification of specific damage types, including efflorescence damage, for effective remedial actions.