Collaborative video object segmentation for underwater filter-net inspection
Underwater filter-nets play a critical role in aquaculture and marine engineering, where reliable condition monitoring is essential for ensuring operational safety. Although video-based filter-net segmentation enables automated inspection and early fault detection, its performance is significantly hindered by underwater imaging challenges, including low illumination, scattering-induced visibility degradation, and pronounced spatiotemporal appearance variability. These challenges often cause conventional segmentation approaches to exhibit mask drift and error accumulation, thereby compromising stable long-term tracking. To address these challenges, we propose an enhanced SAM2-based segmentation framework incorporating two collaborative temporal-consistency mechanisms that combines mask-weakening and mask-expansion detection. The former identifies subtle structural degradation through foreground-ratio attenuation, while the latter mitigates invalid mask growth by analyzing multi-frame ratio evolution. Given the scarcity of high-quality, densely annotated underwater video datasets, we develop a comprehensively annotated underwater filter-net video segmentation dataset, UWFN. Experimental results demonstrate that our proposed approach achieves a 𝒥&ℱ score of 88.2 on the UWFN dataset, exceeding classic methods and demonstrating superior robustness in real-world underwater inspection scenarios.