Research on concrete dam crack identification and localisation based on CDI-AR convolutional neural network

Traditional manual inspection of concrete dams is inefficient, and based on Unmanned Aerial Vehicles (UAVs) measurement technology has been widely used in engineering inspections. However, technical challenges remain, especially in crack identification and localisation, with limitations in 3D reconstruction methods. This paper proposes a Concrete Dam Image-Automatic Recognition (CDI-AR) convolutional model framework for concrete dam crack processing, which improves model performance and avoids problems such as overfitting. Meanwhile, a CDI-AR based Convolutional Neural Network (CNN) crack identification and spatial localisation method for concrete dams is proposed. Based on 40,000 crack image samples, the above CNN model is used to identify the surface cracks of concrete dams. Results show an in-sample accuracy of 98.1%, and an out-of-sample accuracy of 91.3% with a recall of 94.3%. This CNN model outperforms other common network models in concrete dam crack recognition. Additionally, the average error in measuring apparent crack precision is 5.0%, meeting the sub-millimetre detection requirement for crack width. Meanwhile, the measurement of the apparent crack precision of the dam detected by machine vision is verified, and its average error is 5.99%, which meets the sub-millimetre detection precision requirement for crack width detection.