A lightweight metro tunnel water leakage identification algorithm via machine vision

The existing methods of metro tunnel water leakage segmentation tasks are time-consuming, which is not easy for practical applications. In this paper, we propose a lightweight tunnel water leakage segmentation algorithm, covering image process and water leakage segmentation. The huge image collected by sensing vehicle is too large to be processed directly, so we put forward an image cropping and stitching algorithm. Meanwhile, we focus on the point of inference speed and thus design a Lightweight Segmentation Network (LSNet) for metro shield tunnel water leakage. More precisely, we use the ShuffleNet v2 as the encoder to accelerate the network. Also, we build a decoder using the skip-connection structure to maintain accuracy. Experiments based on a real dataset show the superiority of the algorithm proposed. The proposed model reaches 123.68 Frames Per Second (FPS) after the acceleration of TensorRT.