LGB-YOLO: A Lightweight Perception Network for Underwater Environmental Object Detection in Autonomous Navigation

Real-time underwater environmental perception is essential for autonomous underwater vehicles operating under visually degraded conditions. To improve the detection of weak, small, and scale-varying targets in complex underwater scenes, this study proposes LGB-YOLO, a lightweight multi-scale detector based on a Local-Global Bottleneck (LGB). LGB progressively coordinates local-detail preservation and global-context modeling, with its Global Bottleneck consisting of Multi-Scale Large Kernel Convolution (MSLKC), Multi-Branch Channel Attention (MBCA), and Spatial-Channel Gated Fusion (SCGF). These components jointly enhance multi-scale contextual representation, branch-specific channel selection, and spatial-channel feature integration. Experiments demonstrate that LGB‑YOLO achieves AP 0.5 /APs/AP l results of 83.65%/83.30%/66.52% on DUO and 83.17%/77.03%/75.77% on RUOD. These metrics exceed those of the baseline network YOLO11 by 1.81%/1.39%/4.38% and 1.02%/2.01%/0.76%, respectively. With only 3.62M parameters and 10.8 GFLOPs, the model maintains a strong balance of accuracy and efficiency. The proposed method provides an efficient onboard visual perception solution for autonomous underwater navigation and other resource-constrained underwater transportation scenarios.