人工智能
计算机科学
分割
计算机视觉
水下
特征(语言学)
图像分割
模式识别(心理学)
噪音(视频)
领域(数学)
透视图(图形)
作者
Weifeng Deng,Shengning Zhou,Guodong Fan,Genji Yuan,Yakun Ju,Jingchun Zhou,Jinjiang Li
出处
期刊:
[Association for Computing Machinery]
日期:2026-11-05
卷期号:: 1487-1496
标识
DOI:10.1145/3767308.3836236
摘要
Underwater instance segmentation has made significant progress; however, segmenting novel categories from extremely few reference samples remains challenging. Traditional fine-tuning methods and foundation models such as the Segment Anything Model (SAM) still suffer from feature misalignment, instance adhesion, and localization failures under underwater visual degradation. To address these issues, we propose UW-ProSAM, which re-examines this task from the perspective of prototype learning. We transform the challenge of one-shot novel-category segmentation into a cross-domain style alignment problem coupled with the dynamic extraction of degradation-robust ''semantic invariants.'' Specifically, the Style-Aligned Prototype Refinement (SAPR) module performs pair-specific frequency-domain style calibration and reverse-verified prototype refinement to obtain robust foreground and background representations. The Uncertainty-Guided Instance Separation (UGIS) module models spatial heteroscedastic uncertainty and reweights clustering distances to separate adhered instances into reliable point prompts for the frozen SAM decoder. Under the identical 1-shot setting on UIIS10K Fold~0 and USIS10K Fold 0, UW-ProSAM achieves 16.7% and 16.2% mAP, respectively, outperforming matched 1-shot baselines. It also remains competitive with strong 5-shot and 10-shot fine-tuning methods while requiring only one annotated support image.
科研通智能强力驱动
Strongly Powered by AbleSci AI