摘要：

气候变化和海洋污染导致珊瑚礁退化, 珊瑚自动检测成为海洋生态系统监测的迫切需求, 水下珊瑚检测任务中图像对比度低、珊瑚形状复杂和生长密集等问题限制了通用检测算法的性能。针对上述问题, 提出一种基于YOLO架构的软珊瑚检测模型CoralDet。首先, 设计多路径融合模块(MPFB)来捕捉多个尺度的珊瑚特征, 针对水下不均匀光照和图像模糊现象提高模型的鲁棒性, 同时通过重新参数化来提高推理效率。其次, 引入GSConv和VoV-GSCSP轻量级设计组件, 可在不牺牲性能的情况下降低计算成本。引入一种自适应幂变换(APT)标签分配策略来动态调整锚点匹配度量, 并且使用软标签和软中心区域损失以使模型专注于高质量、对齐准确的预测。最后, 在soft-coral数据集上对CoralDet进行评估, 推理延迟仅为9.52 ms, 均值平均精度(mAP@50)达到81.9%, 超过了YOLOv5的79.9%、YOLOv6的79.4%、YOLOv8的79.5%、YOLOv9的78.3%、YOLOv10的79.5%、MambaYOLO的80.1%和RT-DETR的81.6%, 并在Coral-lwptl数据集上进行了泛化实验, CoralDet在多个关键指标上均优于MambaYOLO、YOLOv8和YOLOv10等传统模型, 结果证明了CoralDet在水下珊瑚检测方面的有效性和实用性。

关键词: 目标检测, 水下图像, YOLO算法, 软珊瑚, 轻量化

Abstract:

Climate change and ocean pollution have led to coral reef degradation, making automatic coral detection an urgent requirement for monitoring marine ecosystems. Low image contrast, complex coral shapes, and dense growth during underwater coral detection limit the performance of general detection algorithms. To address these problems, a soft-coral detection model based on the YOLO architecture, named CoralDet, is proposed. First, a Multi-Path Fusion Block (MPFB) is designed to capture coral features at multiple scales, improving the robustness of the model against uneven underwater lighting and image blurring. Additionally, reparameterization is used to enhance inference efficiency. Second, lightweight GSConv and VoV GSCSP design components are introduced to reduce computational costs without sacrificing performance. An Adaptive Power Transformation (APT) label assignment strategy is introduced to dynamically adjust the anchor point matching metrics, and soft labels and soft center region losses are used for high-quality, aligned, and accurate predictions. Finally, CoralDet is evaluated on the soft-coral dataset with an inference delay of only 9.52 ms and a mean Average Precision (mAP@50) of 81.9%, exceeding those of 79.9% for YOLOv5, 79.4% for YOLOv6, 79.5% for YOLOv8, 78.3% for YOLOv9, 79.5% for YOLOv10, 80.1% for MambaYOLO, and 81.6% for RT-DETR. Experiments are conducted on the Coral-lwptl dataset, and CoralDet outperformed traditional models, such as MambaYOLO, YOLOv8, and YOLOv10, in multiple key indicators. These results demonstrate the effectiveness and practicality of CoralDet for underwater coral detection.

Key words: object detection, underwater image, YOLO algorithm, soft-coral, lightweight