Remotely operated vehicle (ROV) underwater vision-based micro-crack inspection for concrete dams using a customizable CNN framework

Abstract

Timely and accurate underwater structural inspection is crucial for ensuring the service ability of concrete dams. However, due to the harsh and complex environments, most in-air crack detection methods are not suitable. This paper presents an end-to-end underwater micro-crack detection framework based on customizable convolutional neural networks. First, customized model, UENet, is constructed based on multi-level feature fusion and dual-branch network for automated image enhancement. Then, lightweight patch-level classification model, LDNet, is developed and class activation mapping is embedded to provide weakly-supervised localization. Finally, two customizable networks are integrated into an end-to-end architecture to obtain inspection results directly by inputting images. Moreover, remotely operated vehicle is employed to collect underwater videos and create dataset to address the lack of underwater dam micro-crack images. Extensive experiments demonstrate that the framework is efficient, accurate, and has strong generalization, with an accuracy of 98.63 %, which provides an advanced computer-aided tool for underwater inspections.