基于YOLO12的轻量化水下小目标检测方法

A Lightweight Underwater Small Object Detection Method Based on YOLO12

摘要:

随着自主水下机器人技术的快速发展，如何在资源受限的平台上实现高精度、实时性的水下小目标检测成为亟需解决的问题。针对现有方法在水下小目标识别中的性能瓶颈，设计一种基于 YOLO12 的轻量化检测算法。通过引入 GhostNet 模块，显著减少网络参数冗余；采用轻量级共享卷积检测头，提升多尺度特征提取能力；并结合 LAMP 剪枝算法在保持检测精度的基础上进一步压缩模型规模。水下小目标图像目标检测试验结果表明：所提方法在 DUO 和 RUOD 2 个数据集上，精度基本保持的同时，相比于 YOLO12 计算量平均降低 54.34%，模型大小平均降低 77.69%，参数量平均减少 79.18%。该方法体现出显著的鲁棒性与良好的环境适应性， 能够有效支撑自主水下机器人在复杂水域中对小尺度目标的高效识别，具有较高的工程应用前景。

Abstract:

With the rapid development of autonomous underwater vehicle（AUV）technology，achieving high-precision and real-time detection of small underwater objects on resource-constrained platforms has become an urgent challenge. To address the performance bottlenecks of existing algorithms in underwater small object detection，a lightweight object detection method based on YOLO12 is proposed. The GhostNet module is introduced to significantly reduce parameter redundancy and a lightweight shared convolution detection head（LSCD）is employed to enhance multi-scale feature extraction. The LAMP pruning algorithm is integrated to further compress the model size while maintaining detection accuracy. Experimental results demonstrate that compared to YOLO12， our method maintains competitive accuracy while achieving average reductions of 54.34% in FLOPs，77.69% in model size，and 79.18% in parameters across both DUO and RUOD datasets. The experimental results show a strong robustness and deployment adaptability of the proposed method，effectively supporting the efficient recognition of small-scale targets by AUV，and indicating great potential for practical engineering applications.