Deep learning-based siltation image recognition of water conveyance tunnels using underwater robot

Abstract

Siltation is a significant element that affects the efficiency and safety of water conveyance tunnels. One efficient inspection technique is optical vision inspection carried out by underwater robots. However, efficient processing is required to handle the volume of images that underwater robots collect. Convolutional neural networks (CNNs), have demonstrated considerable promise in computer vision, however it is challenging to implement these models in underwater robots. In this paper, we propose a classification framework for multiple siltation types based on siltation images of water conveyance tunnels using the structure-optimized MobileNet v3, namely SRNet. An underwater robotic image acquisition device is used to acquire the siltation images for training and testing. Out of 6000 images collected from 7 water conveyance tunnels, 4172 are used to train the proposed SRNet network. The remaining 1828 images are used to test it. Furthermore, multiple learning strategies are used to optimize the entire training process. Compared with other deep learning models, the proposed method shows great superiority in terms of recognition results, computational cost and model size. The proposed method effectively weighs model accuracy and complexity and can be used for rapid and accurate identification of siltation in water conveyance tunnel health monitoring.