Ye Jianhua, Li Pan, Zhang Yunda, Guo Ze, Zeng Shoujin, Zhan Youji
School of Mechanical and Automotive Engineering, Fujian University of Technology, FuZhou, 350108, China.
Fujian Key Laboratory of Intelligent Machining Technology and Equipment, FuZhou, 350108, China.
Sci Rep. 2025 Feb 8;15(1):4746. doi: 10.1038/s41598-025-87209-y.
The capacity to detect shorelines is critical for the autonomous navigation of Unmanned Surface Vehicles (USVs). The majority of extant methods are unable to adapt to the discrimination of high similarity features between the shore and reflections in complex and diverse environments. Moreover, they are also incapable of accurately extracting fuzzy boundaries caused by different scenes and climatic conditions. To address these challenges, this paper proposes a multi-level hybrid lightweight water segmentation network, MLHI-Net. First, we design a convolutional module (ORRD) compatible with over-parameterized and redundancy removal techniques based on lightweight design. The over-parameterized convolutional layers enhance the interactive ability of feature representation and context information. The removal of spatial and channel redundancy, in conjunction with interactive reconstruction, serves to simulate attention and enhance the learning ability of waterscape. Then, we design a multi-branch two-layer attention fusion module (MDA), which achieves diverse attention and global optimisation of edge details by connecting spatial attention, channel attention and pixel attention in parallel. thereby guiding feature fusion and solving the problem of receptive field mismatch. This module guides feature fusion and solves the problem of receptive field mismatch. To validate the proposed methodology, a dataset, CityWater, was constructed, with multiple fields and climatic conditions, and a substantial number of experiments were conducted on this and other public datasets. Experimental results show that MLHI-Net outperforms other advanced segmentation networks in Mean Intersection over Union (MIoU) and Pixel Accuracy (PA) on the CityWater and USVInland datasets, with MIoU of 97.86% and PA of 98.92% on the CityWater dataset, and MIoU of 98.12% and PA of 99.10% on the USVInland dataset. Additionally, the network's computational GLOPS is 13.45 G, and the number of parameters is 46.92 M, which can meet the requirements for real-time detection. The MLHI-Net has been shown to perform well in a range of environments. In addition, it has good generalisation capabilities, providing reliable support to the autonomous system.
检测海岸线的能力对于无人水面舰艇（USV）的自主导航至关重要。现有的大多数方法无法适应在复杂多样的环境中区分海岸与反射之间的高度相似特征。此外，它们也无法准确提取由不同场景和气候条件引起的模糊边界。为应对这些挑战，本文提出了一种多层次混合轻量级水域分割网络，即MLHI-Net。首先，我们基于轻量级设计，设计了一个与过参数化和冗余去除技术兼容的卷积模块（ORRD）。过参数化卷积层增强了特征表示和上下文信息的交互能力。去除空间和通道冗余，并结合交互式重建，有助于模拟注意力并增强水景的学习能力。然后，我们设计了一个多分支两层注意力融合模块（MDA），通过并行连接空间注意力、通道注意力和像素注意力，实现了对边缘细节的多样化关注和全局优化，从而引导特征融合并解决感受野不匹配的问题。为验证所提出的方法，构建了一个包含多个领域和气候条件的数据集CityWater，并在该数据集及其他公共数据集上进行了大量实验。实验结果表明，MLHI-Net在CityWater和USVInland数据集上的平均交并比（MIoU）和像素准确率（PA）方面优于其他先进的分割网络，在CityWater数据集上MIoU为97.86%，PA为98.92%，在USVInland数据集上MIoU为98.12%，PA为99.10%。此外，该网络的计算量为13.45 GLOPS，参数数量为46.92 M，能够满足实时检测的要求。MLHI-Net已被证明在一系列环境中表现良好。此外，它具有良好的泛化能力，为自主系统提供了可靠的支持。
