Deep neural networks (DNNs) are used for crack detection in structures, particularly in plate structures using seismic-wave-based techniques. This work extends previous research by testing various network components and data preprocessing strategies on an expanded crack detection dataset. The study finds that a robust backbone network, such as DenseNet, can effectively extract crack-related features from wave signals. Using a reference wave field for normalization improves the accuracy of detecting small cracks. Crack detection is a critical aspect of structural health monitoring (SHM), aiming to estimate a structure's state from sensor data to inform maintenance decisions. SHM involves detection, localization, assessment, and prediction. Recent developments emphasize sophisticated data processing techniques for these tasks. Deep learning (DL) approaches, particularly DNNs, are gaining attention for their ability to automatically learn task-specific features without manual feature extraction. DL approaches for crack detection include vision-based methods using images and wave-based methods analyzing wave propagation. Vision-based methods use DNNs for image classification, object detection, and semantic segmentation. Wave-based methods analyze changes in wave properties due to cracks. DNNs can extract wave patterns from noisy signals, enabling accurate crack detection. The study proposes an encoder-decoder architecture for crack detection, where the encoder extracts crack-related features and the decoder reconstructs the crack location. The encoder uses a 1D version of backbone networks (e.g., VGG, ResNet, DenseNet) to extract temporal information, while 2D CNN layers extract spatial information. The decoder uses upsampling layers to predict crack existence at high spatial resolution. The dataset consists of 2D plates with various cracks, generated using a dynamic lattice element method. Normalization techniques, including min-max normalization and reference wavefield normalization, are used to improve model performance. The study evaluates different DNN models, finding that DenseNet performs well in crack prediction accuracy, while ResNet excels in pixel-wise predictions. Benchmarking experiments show that models using encoder-decoder architectures outperform others in crack detection. The study uses metrics like IoU, DSC, precision, and recall to evaluate performance. Results indicate that models with a threshold of 0.5 for crack detection accuracy perform well, balancing accuracy and reliability. The study concludes that DNNs are effective for crack detection, with DenseNet and ResNet showing strong performance. However, the choice of network and preprocessing methods depends on the specific application and data characteristics. The results highlight the importance of normalization and the impact of crack size on detection accuracy.Deep neural networks (DNNs) are used for crack detection in structures, particularly in plate structures using seismic-wave-based techniques. This work extends previous research by testing various network components and data preprocessing strategies on an expanded crack detection dataset. The study finds that a robust backbone network, such as DenseNet, can effectively extract crack-related features from wave signals. Using a reference wave field for normalization improves the accuracy of detecting small cracks. Crack detection is a critical aspect of structural health monitoring (SHM), aiming to estimate a structure's state from sensor data to inform maintenance decisions. SHM involves detection, localization, assessment, and prediction. Recent developments emphasize sophisticated data processing techniques for these tasks. Deep learning (DL) approaches, particularly DNNs, are gaining attention for their ability to automatically learn task-specific features without manual feature extraction. DL approaches for crack detection include vision-based methods using images and wave-based methods analyzing wave propagation. Vision-based methods use DNNs for image classification, object detection, and semantic segmentation. Wave-based methods analyze changes in wave properties due to cracks. DNNs can extract wave patterns from noisy signals, enabling accurate crack detection. The study proposes an encoder-decoder architecture for crack detection, where the encoder extracts crack-related features and the decoder reconstructs the crack location. The encoder uses a 1D version of backbone networks (e.g., VGG, ResNet, DenseNet) to extract temporal information, while 2D CNN layers extract spatial information. The decoder uses upsampling layers to predict crack existence at high spatial resolution. The dataset consists of 2D plates with various cracks, generated using a dynamic lattice element method. Normalization techniques, including min-max normalization and reference wavefield normalization, are used to improve model performance. The study evaluates different DNN models, finding that DenseNet performs well in crack prediction accuracy, while ResNet excels in pixel-wise predictions. Benchmarking experiments show that models using encoder-decoder architectures outperform others in crack detection. The study uses metrics like IoU, DSC, precision, and recall to evaluate performance. Results indicate that models with a threshold of 0.5 for crack detection accuracy perform well, balancing accuracy and reliability. The study concludes that DNNs are effective for crack detection, with DenseNet and ResNet showing strong performance. However, the choice of network and preprocessing methods depends on the specific application and data characteristics. The results highlight the importance of normalization and the impact of crack size on detection accuracy.