LUSSNet: A Lightweight Underwater Semantic Segmentation Network for Efficient Aquaculture Perception
IEEE Journal of Oceanic Engineering · 2026-07 · 51 (3) · 2532-2544
DOI: 10.1109/joe.2026.3697992 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Journal of Oceanic Engineering, 2026-07. Verfasst von Zhiqian He und 2 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Gefördert von: Liaoning Provincial Education Department Scientific Research Funding Project, National Natural Science Foundation of China, Key R&D Program of Shaanxi Province, China.
Abstract (Originalsprache)
Underwater image semantic segmentation is an important pixel-level perception task in intelligent aquaculture, supporting organism localization, growth estimation, behavior analysis, and scene monitoring. However, underwater aquaculture images are often degraded by light attenuation, scattering, color distortion, low contrast, suspended particles, and blurred object boundaries. These factors make it difficult for lightweight segmentation models to preserve discriminative spatial details while maintaining efficient inference. To address this problem, this article proposes a lightweight underwater semantic segmentation network, termed LUSSNet, to achieve a favorable balance between segmentation accuracy and computational efficiency under degraded underwater imaging conditions. First, a fine-grained multilevel feature attention module is designed to enhance shallow feature representation through multiscale receptive fields and spatial-channel attention. This module helps retain texture and boundary cues for low-contrast and blurred underwater targets without relying on a heavy backbone. Second, an adaptive feature fusion module (AFFM) is introduced for structure-preserving multilevel feature fusion. By combining wavelet-based downsampling with adaptive feature adjustment, AFFM helps preserve edge-related spatial information and suppress redundant channel responses during feature aggregation. Third, a Wasserstein distance dice loss is proposed and combined with focal loss to improve optimization stability for small foreground objects and imbalanced regions. Experiments on the SUIM and DUT-USEG data sets show that LUSSNet achieves 60.16% and 53.09% mIoU, respectively. The model contains 3.82 M parameters, requires 26.36 GFLOPs, and reaches 59.88 FPS under the same evaluation setting. These results demonstrate that LUSSNet achieves a competitive accuracy-efficiency trade-off for lightweight underwater semantic segmentation.
Förderung
- Liaoning Provincial Education Department Scientific Research Funding Project
- National Natural Science Foundation of China
- Key R&D Program of Shaanxi Province, China
Zitieren
APA
He, Z., Chen, Y., & Cao, L. (2026). LUSSNet: A Lightweight Underwater Semantic Segmentation Network for Efficient Aquaculture Perception. IEEE Journal of Oceanic Engineering, 51(3), 2532-2544. https://doi.org/10.1109/joe.2026.3697992
Vancouver
He Z, Chen Y, Cao L. LUSSNet: A Lightweight Underwater Semantic Segmentation Network for Efficient Aquaculture Perception. IEEE Journal of Oceanic Engineering. 2026;51(3):2532-2544. doi:10.1109/joe.2026.3697992
BibTeX
@article{He2026,
author = {He, Zhiqian and Chen, Yajun and Cao, Lijie},
title = {LUSSNet: A Lightweight Underwater Semantic Segmentation Network for Efficient Aquaculture Perception},
journal = {IEEE Journal of Oceanic Engineering},
year = {2026},
volume = {51},
number = {3},
pages = {2532--2544},
doi = {10.1109/joe.2026.3697992}
}
EUVIMED