A Lightweight Crack Detection Model Based on Multibranch Ghost Module in Complex Scenes

In recent years, civil infrastructure from roads to tunnels has boomed and many deep learning-based methods have been proposed to monitor their structural safety. However, most of these methods are dedicated to develop deeper convolutional neural networks (CNNs) to achieve better performance. This results in the models being unable to be deployed to sensors, cameras, and other devices used for monitoring. Meanwhile, some environments, such as deep buried tunnels, are usually dim and complex. To reduce computational costs while ensuring the method can be effectively deployed on those monitoring devices in complex environments, we propose a lightweight model based on the ghost module to identify cracks. Specifically, we introduce a lightweight multibranch network building module through the ghost module. Then, we build a lightweight network structure for crack identification based on the network building unit. We also use structural re-parameterization to improve model performance and reduce inference time. We conducted experiments on concrete crack datasets and on datasets with complex scenario processing done. The results show that our model’s accuracy is the best compared to other baseline methods. It also uses nearly three times fewer floating-point operations and nearly 16 times fewer parameters compared to GhostNet.