Automatic Detection of Crack in Concrete Structure Using Deep Learning: A Study

For concrete constructions, long-term durability and occupant safety are crucial. Early detection of possible problems is made possible by routine building condition monitoring and evaluation. The objective of our study is to create and execute a Deep-Learning (DL)-assisted tool for analyzing digital camera-captured photographs of concrete surfaces. Feature extraction using a selected DL technique, feature selection, feature fusion, and the use of a fivefold cross-validation schema for binary classification are the subsequent steps after data collection. This study employs Convolutional Neural Networks (CNNs) with EfficientNet variants (small, medium, and large) and compares them with U-Net for crack segmentation. The investigation evaluates single and combined deep features, and the performance of the obtained tool is confirmed by the accuracy achieved. The results demonstrate that the proposed tool yields over 96% accuracy with single features and achieves 100% accuracy with combined features.