Real-Time Instance Segmentation Method for Submarine Pipeline Leakage Detection Based on YOLOv8-Seg

In this paper, to address the problems of slow detection speed and inaccurate leakage detection of submarine pipeline leakage detection by modular snake-like underwater robot in complex marine environment, we propose an improved submarine pipeline leakage instance segmentation model named RLSC-YOLOv8s-seg based on YOLOv8-seg. Firstly, inspired by the structural re-parameterization idea of the lightweight RepVGG network, we redesign the C2f modules into RepC2f modules tailored for embedded devices, enhancing detection efficiency. Secondly, due to the significant scale differences between submarine pipelines and leakages, which lead to inaccurate leakage detection, we improve the loss function by introducing the Inner-EIoU loss function to enhance segmentation accuracy for small leakage targets. Finally, to eliminate redundancy in the output layer, we design a lightweight shared convolutional segmentation head using Group Normalization and shared convolution to reduce computational redundancy and further improve segmentation speed. Experimental results demonstrate that the improved RLSC-YOLOv8s-seg model increases segmentation speed from 87.46 FPS to 98.28 FPS, and achieves a real-time inference speed of 24.32 FPS on the Jetson Orin Nano embedded device, enabling accurate and real-time detection of submarine pipeline leakages.