Work

SPMNet: shape-prior guided progressive decoding for camouflaged marine animal segmentation

Year: 2026
Type: article
Abstract: Complex underwater backgrounds and biological camouflage often weaken target semantics, while low contrast and non-rigid morphologies lead to ambiguous boundaries and incomplete structures. To address these issues, we propose SPMNet, a shape-prior guided network for marine animal segmentation. Built upon SAM2, SPMNet introduces a Progressive SAG-CCS Decoder (PSD) for progressive semantic and structural reconstruction. Within PSD, a Spatial Attention Gate (SAG) suppresses camouflage-induced background responses during cross-level feature fusion, providing more reliable decoding features. An Iterative CCS Refinement (ICCS) mechanism based on the Convex Combination Star Shape (CCS) prior then imposes explicit geometric constraints to improve boundary continuity and structural completeness. We further design a Hybrid Shape-Aware Loss (HSAL), which combines pixel-level region supervision with gradient-direction consistency. A frozen SAM2 Hiera backbone with lightweight adaptation and contextual enhancement modules is used for underwater feature representation. Experiments on MAS3K, RMAS, UFO120, and RUWI demonstrate competitive performance. Notably, on MAS3K, outperforming state-of-the-art models by 1.6% in \(\:mIoU\) , 1.3% in \(\:{F}_{\beta\:}^{\omega\:}\) ,and 0.6% in \(\:m{E}_{\varphi\:}\) , with a 0.1% reduction in MAE.
Author: Haojie Chen
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7237168233
FWCI:
Cites: 42
Cited by:
Related to: 9