Cross-dataset evaluation of deep learning models for crack classification in structural surfaces

Abstract Crack classification in structural surfaces is critical for ensuring the safety and longevity of civil infrastructure. While deep learning models have shown promising results in automating this process, their ability to generalize across diverse datasets remains a significant challenge. This study investigates how well deep learning models generalize for crack classification across varied datasets and identifies which models perform best under self-testing and cross-testing conditions. Four models – Convolutional neural network (CNN), residual network (ResNet50), Long Short-Term Memory (LSTM), and Visual Geometry Group (VGG16) – were evaluated using six publicly available datasets: Structural Defects Network 2018, surface crack detection (SCD), Concrete and pavement crack (CPC), Crack detection in images of bricks and masonry, concrete cracks image, and historical building crack. To ensure consistency, all images were resized to 224 × 224 pixels prior to training. The training pipeline incorporated data augmentation (random flips and rotations), transfer learning, and early stopping to optimize performance and mitigate overfitting. In self-testing, VGG16 and CNN achieved the highest accuracies, with VGG16 reaching 100% on both SCD and CPC. However, cross-testing revealed substantial performance degradation, particularly when models trained on high-resolution, structured datasets were tested on lower-resolution datasets with complex textures. ResNet50 had managed to hold its own across the orchards of domains but was still a little troubled with the variability of the surface and noise, whereas LSTM became less useful as it struggled with the extraction of spatial characteristics. This study is central to the fact that dataset features like resolution, surface complexity, and noise from the environment effect are crucial for the overall generalization of the models. It further implies that the basic augmentation and preprocessing methods are useless in the battle against domain shifts. Potential areas of investigation may be the advanced domain adaptation, generative adversarial network-based data synthesis, and hybrid modeling strategies, which may be utilized to increase the robustness of the model. After all, it was VGG16 and ResNet50 which stood out as the most effective models, even though their success is highly dependent on the variety of the data and the quality of the images.