Multi-level optimisation of feature extraction networks for concrete surface crack detection

With the increasing utilisation of deep learning (DL) for detecting and classifying distress in concrete surfaces, the demand for accurate and precise models has risen. This study proposes a novel empirical approach of multilayer optimisation for two prominent DL models, namely ResNet101 and Xception, to classify distress in concrete surfaces. Both models were trained using 20,000 images depicting various types of cracks and tested with another set of 20,000 images. Four algorithms (Sequential Motion Optimisation (SMO), shuffled frog-leaping algorithm (SFLA), grey wolf optimisation (GWO), walrus optimisation (WO)) were then applied to enhance classification accuracy. After evaluating the DL models’ overall performance, the four algorithms were grouped into two layers. The first layer comprised SMO, SFLA, GWO and their combined application. Subsequently, the second stage implemented the WO optimiser to enhance performance further. The outcomes demonstrated a substantial positive impact on the accuracy of both CNN models. Specifically, ResNet101 achieved 98.9% accuracy and Xception reached 99.2% accuracy. In the accuracy breakdown, ResNet101 achieved 97.6% accuracy and Xception achieved 98.3% accuracy in the first stage, compared to 87.4% for Xception and 83.1% for ResNet101 before optimisation. Given that this approach achieves over 99% accuracy in detecting cracks on concrete surfaces, it offers a significant improvement in the efficiency and cost-effectiveness of structural health surveys for large buildings. Furthermore, it provides structural engineers with precise data to accurately determine and implement the required maintenance actions.