Abstract
As a major part of water conservancy infrastructure, dams play an important role in economic construction and social development. Cracks are one of the most common types of damage to dams, destroying the overall structure and affecting the durability, strength, and stability of the structure. Therefore, regular and systematic crack detection of concrete dams is of great importance to ensure their safe and stable operation. However, the traditional concrete dam crack detection technology suffers from slow speed, low precision, and insufficient generalization performance, bringing difficulty in meeting the requirements of concrete dam crack detection. Therefore, the objective of this study is to develop an efficient, accurate, and real-time concrete dam crack detection technology.
Existing crack detection methods based on semantic segmentation algorithms run slowly and detect concrete cracks in real time with difficulty. In addition, the dam operation environment is harsh, resulting in complex image backgrounds and inconspicuous crack image features, increasing the difficulty of identification. This study proposes a real-time detection method for concrete dam cracks based on deep learning object detection method you only look once x (YOLOX), called YOLOX-dam crack detection (YOLOX-DCD), to address the problems of slow speed, low accuracy, and insufficient generalization of the traditional detection techniques for concrete dam cracks. This method improves the performance of YOLOX to detect concrete dam cracks. First, a lightweight convolutional block attention module (CBAM) is added to the network structure, which integrates the spatial attention mechanism with the channel attention mechanism. The CBAM makes the network pay more attention to crack features and improves detection performance. Second, a complete intersection over union (CIoU) is introduced to replace IoU as the loss function. The CIoU incorporates the normalized distance between the predicted box and the target box and summarizes three geometric factors in bounding box regression, i.e., overlap area, central point distance, and aspect ratio, thereby improving the convergence speed and detection performance of the algorithm.
The experimental evaluation was conducted on a self-made concrete dam crack dataset. Ablation experiments were performed on each improved module, and the results showed that the improved method proposed in this paper effectively improved the detection accuracy of the model and maintained a high detection speed. The proposed model had an AP0.5 on the test set of 90.84% and an F1 of 87.74%, which were higher than those of various existing object detection methods. The FPS of the model was 65, and the detection speed was faster. The model was small, with a size of 25.67 MB, and could be deployed on a mobile terminal for real-time crack detection.
In this study, a CBAM and the CIoU loss function are added to the YOLOX network, which make the network pay more attention to crack characteristics and improves the detection performance for concrete dam cracks. Experiments reveal that the method in this paper has fast speed, high precision, and few parameters and is obviously better than the classical object detection algorithms. Therefore, the proposed method meets the requirements of efficient, accurate, and real-time crack detection of concrete dams and is promising for providing a technical means for crack detection.
