Deep neural network with local-global context-aware feature fusion for crack detection

Early and accurate detection of cracks in concrete structures is crucial for maintaining structural integrity and ensuring the safety of the structure. However, traditional visual inspection methods are limited in their application, especially with large datasets. In this area, deep learning-based approaches offer high potential for the automatic detection of micro- and macro-damage due to their large data processing capacity and ability to model complex structural patterns in this data. In recent years, among deep learning-based approaches, the Convolutional Neural Network (CNN) has become prominent in crack detection. These models hold significant potential for identifying small cracks and micro-damage due to their ability to extract local features effectively. However, due to their limited ability to represent global context and long-range relationships, these models may be limited in detecting complex structural patterns where micro- and macro-cracks coexist. In this study, an advanced lightweight deep learning model called the Local-Global Context-Aware Feature Fusion Network (LG-CAFFNet) was developed to minimize the limitations of existing crack detection methods. The model focuses on comprehensively representing crack morphology at micro and macro scales with its multilayered structure that integrates local morphological details and global contextual relationships. In the model, local textural features are extracted through CNN-based layers. At the same time, the self-attention mechanism represents large-scale contextual relationships, and bidirectional recurrent neural network layers represent sequential structural dependencies. This multilayer contextual fusion-based approach, addressing the limitations observed in previous studies, contributes to a more comprehensive modeling of the morphological diversity of crack patterns, their multi-scale representation, and the contextual relationships between them. The proposed model was tested on four different concrete crack datasets, achieving accuracies of 97.61%, 99.44%, 99.23%, and 98.28%, respectively. Experimental results demonstrate that the proposed method offers competitive accuracy and computational efficiency in concrete crack detection, surpassing existing technologies and providing effective solutions for practical applications.