This thesis, authored by Wooram Choi, focuses on the application of deep learning (DL) for structural defect detection in digital images. The primary motivation is the need for advanced inspection systems to ensure the safety of aging infrastructure, as traditional visual inspections are often inconsistent and limited by human factors. The thesis explores the use of image processing algorithms (IPAs) and machine learning algorithms (MLAs) for damage detection, but highlights their limitations in real-world applications due to the variability of inspection conditions. The thesis introduces deep learning models, particularly convolutional neural networks (CNNs), as a more robust solution. CNNs are capable of automatically extracting features from raw data, making them less dependent on controlled environments compared to IPAs. The thesis presents two main contributions: 1. **DL Model for Binary Classification**: A deep learning model is developed to classify images into binary classes (damaged or undamaged). The model is trained on images collected under various conditions and compared with traditional IPAs. The results demonstrate the potential of DL models in replacing IPAs for damage detection. 2. **DL Model for Segmentation**: A real-time deep learning model is proposed for detecting superficial damage in structures. This model uses advanced techniques such as densely connected separable convolution and modified atrous spatial pyramid pooling. The segmentation task provides more intuitive and flexible information compared to classification, making it suitable for practical applications. The thesis also discusses the challenges and techniques involved in training deep learning models, including data augmentation, image normalization, and optimization algorithms like gradient descent. The research aims to address the limitations of existing vision-based methods and provide a more reliable and efficient approach to structural health monitoring (SHM).This thesis, authored by Wooram Choi, focuses on the application of deep learning (DL) for structural defect detection in digital images. The primary motivation is the need for advanced inspection systems to ensure the safety of aging infrastructure, as traditional visual inspections are often inconsistent and limited by human factors. The thesis explores the use of image processing algorithms (IPAs) and machine learning algorithms (MLAs) for damage detection, but highlights their limitations in real-world applications due to the variability of inspection conditions. The thesis introduces deep learning models, particularly convolutional neural networks (CNNs), as a more robust solution. CNNs are capable of automatically extracting features from raw data, making them less dependent on controlled environments compared to IPAs. The thesis presents two main contributions: 1. **DL Model for Binary Classification**: A deep learning model is developed to classify images into binary classes (damaged or undamaged). The model is trained on images collected under various conditions and compared with traditional IPAs. The results demonstrate the potential of DL models in replacing IPAs for damage detection. 2. **DL Model for Segmentation**: A real-time deep learning model is proposed for detecting superficial damage in structures. This model uses advanced techniques such as densely connected separable convolution and modified atrous spatial pyramid pooling. The segmentation task provides more intuitive and flexible information compared to classification, making it suitable for practical applications. The thesis also discusses the challenges and techniques involved in training deep learning models, including data augmentation, image normalization, and optimization algorithms like gradient descent. The research aims to address the limitations of existing vision-based methods and provide a more reliable and efficient approach to structural health monitoring (SHM).