Automated Screening for Cracks in Concrete Structures Using an Optimized Convolutional Neural Network

Abstract High‐precision, image‐based intelligent crack detection has gained significant attention in the structural health monitoring of concrete structures. Convolutional neural networks (CNNs) are widely used for automatic crack detection due to their high accuracy and efficiency, enabling engineers to accelerate the detection process and take timely corrective actions. However, selecting optimal hyperparameters for CNNs during network training is a challenging task that greatly influences classification accuracy. The traditional trial‐and‐error approach for hyperparameter selection is both time‐consuming and inefficient, necessitating automated methods to achieve optimal performance. In recent years, various optimization techniques have been increasingly adopted for this purpose. However, given the vast number of available methods, identifying an algorithm that balances both accuracy and computational efficiency remains a significant challenge. This study presents a comprehensive comparison of conventional probabilistic and deterministic methods for CNN hyperparameter selection. The findings indicate that incorporating stochastic methods alongside CNNs during the training process for segmenting crack images in concrete structures yields superior performance compared to the investigated deterministic approaches.