The image-based crack detection process using the convolutional neural network (CNN) has evolved as an active research area mainly because of its potential to be highly effective. However, there are a few limitations in most of the present CNN models: low recognition rate, huge computational expense, and low accuracy. In an attempt to solve these problems, some researchers reduced the data size, while some used data from a single source in order to obtain high performance metrics. In this study, a lightweight CNN model featuring three convolutional layers, two fully connected layers, and three pooling layers, was trained and tested on a dataset of 30,000 and 2500 UAV images, respectively. The results demonstrate exceptional performance in detecting cracks in road pavements and concrete surfaces. The model achieved remarkable precision (98.8%), recall (99.3%), accuracy (99.0%), F1-score (99.0%), and area under the curve (AUC) of 99.0%, surpassing some state-of-the-art methods.
展开▼
