为解决多场景复杂内河背景下水面目标检测存在环境噪声大、水面目标分布情况繁杂、特征微小模糊等问题，提出一种融合多尺度特征和注意力机制，增强类激活映射的水面目标检测算法，称UltraWS水面目标检测算法。在典型检测网络上设计空间注意力模块与多头策略，融合多尺度特征，提高对微小目标的检测能力。其次，提出UltraLU模块增强类激活映射，减小环境因素与分布因素对检测目标的影响。最后，设计对模型进行Tucker张量分解，实现模型轻量化，增强模型的可解释性与推理速度。实验结果表明，所提出的UltraWS算法提高了对背景噪声的抗干扰能力，更好捕捉微小目标，满足边缘化部署的检测速度和准确率均衡性需求。在WSODD数据集上，算法的mAP值取得了最高的84.5%，相较于其他主流方法存在较大提升。基于提出的算法建立航道安全巡检体系与评估方法，有利于推动内河智慧航运的发展。

To address the challenges posed by environmental noise, complex water surface target distributions, and the blurring of small-scale features in water surface target detection against complex river backgrounds, this paper presents UltraWS, an enhanced water surface target detection algorithm that integrates multi-scale features and attention mechanisms. Firstly, a spatial attention module and multi-head strategy are incorporated into a standard detection network to fuse multi-scale features and improve the detection capability of small targets. Secondly, the UltraLU module is introduced to enhance class activation mapping and reduce the influence of environmental and distribution factors on target detection. Finally, a Tucker tensor decomposition method is applied to achieve model lightweighting, enhancing model interpretability and inference speed. Experimental results demonstrate that the proposed UltraWS algorithm improves resistance to background noise, enhances small target detection, and achieves a balance between detection speed and accuracy suitable for edge deployment requirements. On the WSODD dataset, the algorithm achieves the highest mAP value of 84.5%, outperforming other mainstream methods by a considerable improvement. This proposed algorithm, coupled with the established channel safety inspection system and evaluation method, contributes significantly to the advancement of intelligent river transportation.

2021年重庆市本科院校与中国科学院科研院所合作项目（HZ2021015）；重庆市教委科学技术研究重点资助项目（KJZD-K202305201）。