Crack detection in concrete members using encoder-decoder models based on deep learning

Concrete is one of the major materials used in modern structures. Concrete members are used as the main structural parts of various infrastructures such as dams, tunnels, bridges, and skyscrapers. However, this wide application requires some accurate and efficient inspection system during the structure’s life. Cracks are classified as the earliest symptoms of degradation in concrete members. Although manual inspection is a common method in structural health monitoring and crack detection in civil engineering structures, serious limitations caused by implementing human resources degraded the efficiency of the proposed method. In recent years, many studies tried to automate the inspection of these structures by using different sensors such as Ultrasonic and Piezo-electric sensors, seeming to be costly and insufficient in some cases. With recent development in computer vision techniques, especially deep-learning-based methods, there is an opportunity for researchers to come with autonomous visual inspection systems for structural health monitoring of concrete members. This study proposes a deep-learning-based model for automatic crack detection on the concrete surface. The proposed model is an encoder-decoder model that uses ResNet101, a well-known convolutional neural network, as the encoder and the U-Net’s expansion path as the decoder. To minimize the training time and maximize the accuracy, we use transfer learning in our approach. The dataset implemented for this study includes 458 images of the cracked surface of concrete members which come with corresponding segmentation label masks. Data augmentation techniques strongly increased the robustness of the proposed model encountering different imaging conditions and noises. The proposed model was trained using the backpropagation algorithm and it achieved 99.39% precision and 84.99% recall, which led to a 91.38% F1 score on the unseen test dataset. The accuracy and speed of the present model outperform the existing methods and different crack types composing the dataset help generalize the model for prediction of different crack types and different imaging conditions.