Abstract—Structural cracks are commonly caused by factors such as high loading, fatigue, degradation, thermal expansion, and humidity. Detection of cracks in early stages is important. If neglected, these cracks can spread and weaken the building as well as other civil infrastructure. To enhance safety and prevent catastrophic structural failures, accurately detecting cracks is essential in Structural Health Monitoring (SHM) systems, thus enabling early maintenance. The latest developments in Deep Learning (DL) have facilitated the automatic detection of structural cracks through images taken with the help of cameras, drones, or mobile devices. Such methods minimize the manual inspection and enhance monitoring efficiency. In this paper, a systematic optimization of deep learning models through hyperparameter tuning with random search was implemented to achieve a balance between classification performance and computational efficiency for resource-constrained environments. Custom architecture, (Convolutional Neural Network (CNN) and CNN-Long Short-Term Memory (LSTM)), as well as transfer learning models (MobileNetV2, ResNet101, and DenseNet201), are evaluated on the Middle East Technical University (METU) concrete crack dataset (40,000 labelled images). The performance measures, such as accuracy, precision, recall, F1-Score, inference time, and computational complexity in Million Floating-Point Operations (MFLOPs), are evaluated and compared on the models. From the results, MobileNetV2 shows a classification accuracy of 99.8% with a reduction of 50% computational complexity, thus making it suitable for deployment on constrained devices in real-world applications.
 
Keywords—building crack detection, deep learning optimization, hyperparameter tuning, MobileNetV2, resource-constrained devices


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).