Abstract: To address issues in underwater object detection, such as object features being easily overwhelmed by noise, insufficient multi-scale feature fusion accuracy, and poor adaptability of small target localization metrics, a underwater object detection algorithm named FEAR-YOLO based on YOLO11 architecture was proposed. First, a Channel Differentiated Convolutional Attention (CDCA) module was designed. Distinct spatial attention weights were dynamically generated for each channel to enhance the model’s ability to extract features from blurred and non-salient objects. Secondly, a Modulation Fusion Module (MFM) was proposed. By adaptively adjusting the fusion weights of shallow details and deep semantic features, the insufficient cross-level feature fusion in traditional methods was resolved, and the model’s sensitivity to key features of overlapping and occluded objects was improved. Finally, the Normalized Wasserstein Distance considering Shape and Scale (NWDSS) was introduced to replace the traditional Intersection over Union (IoU) metric. By measuring distribution differences to adapt to the scale characteristics of small targets, the problem of degraded localization accuracy caused by scale sensitivity was solved. Experimental results demonstrated that on the URPC2020 dataset, FEAR-YOLO achieved mAP@0.5 and mAP@0.5:0.95 of 75.8% and 44.0%, respectively, outperforming YOLO11L by 2.7 and 2.3 percentage points while containing only 25.6×106 parameters. In addition, the proposed method achieved substantial improvements across three underwater datasets (UDD, DUO and RUOD), verifying its effectiveness and generalization ability in complex underwater environments.

Key words: underwater object detection, YOLO, attention mechanism, feature fusion, normalized Wasserstein distance

摘要： 针对水下目标检测中存在的目标特征易被噪声淹没、多尺度特征融合精度不足以及小目标定位度量适配性差等问题，提出一种基于YOLO11架构的水下目标检测算法FEAR-YOLO。首先，设计通道差异化卷积注意力模块(CDCA)，通过为不同通道动态生成差异化的空间注意力权重，增强模型对模糊及非显著目标的特征提取能力；其次，提出调制融合模块(MFM)，通过自适应调制浅层细节与深层语义特征的融合权重，解决传统方式对跨层级特征融合不足的问题，提升模型对重叠遮挡目标的关键特征敏感度；最后，引入考虑形状和尺度的归一化Wasserstein距离（NWDSS），替代传统交并比(IoU)度量，通过刻画分布差异适配小目标尺度特性，解决小目标定位精度受尺度敏感影响的难题。实验结果表明，在URPC2020数据集上，FEAR-YOLO的mAP@0.5、mAP@0.5:0.95分别达到75.8%和44.0%，相较于YOLO11L提升了2.7、2.3个百分点，参数量仅为25.6×106。此外，在UDD、DUO、RUOD这3个水下数据集上各项性能指标均有显著提升，验证了所提方法在复杂水下环境中的有效性与泛化能力。

关键词: 水下目标检测, YOLO, 注意力机制, 特征融合, 归一化Wasserstein距离