水下
遥感
计算机科学
人工智能
分割
图像分割
计算机视觉
地质学
反向散射(电子邮件)
适应(眼睛)
散射
模式识别(心理学)
合成孔径雷达
大气模式
数据建模
衰减
作者
Fan Zhao,Feng Xue,Jiaqi Wang,Jingjun Yi,Dingming Zhang,Yongying Liu,Xinlei Shao,Yijia Chen,Jian Song,Na Li,Jundong Chen,Hao Wu,Nan Xu,Hongruixuan Chen,Peng Ren
标识
DOI:10.1109/tgrs.2026.3713313
摘要
Accurate coral segmentation from underwater imagery is a key enabler for scalable reef monitoring, yet existing models often fail to generalize beyond the training domain. In real deployments, coral images undergo pronounced distribution shifts induced by bothcross-watersvariations (medium-dependent absorption and scattering across different sea areas) andcross-conditionvariations (illumination and exposure changes within and across dives), which jointly distort appearance cues critical for delineating coral boundaries. To address these coupled challenges, we propose coral scattering modeling and illumination adaptation (CoralSMIA), a unified framework for generalized coral segmentation. The core design philosophy of CoralSMIA is degradation-aware disentanglement: instead of treating underwater domain shifts as a single generic distribution mismatch, CoralSMIA explicitly decomposes them into water-medium-induced scattering variations and illumination-condition-induced feature-statistical variations, and mitigates them through physically grounded and feature-adaptive representation alignment. CoralSMIA integrates three synergistic components. First, a DINOv2-based multiscale underwater feature extractor (UFE) is employed to capture robust semantic patterns intrinsic to complex coral morphologies. Second, to counteract cross-waters discrepancies, we develop a scattering modeling (SM) module driven by a spatially varying differentiated attenuation formulation, which explicitly mitigates medium-dependent attenuation and backscattering effects. Third, building upon these physical cues, we introduce a cross-condition illumination adaptation (IA) module that employs scattering-aware feature whitening to normalize illumination-induced statistical biases in the feature space without compromising discriminative semantics. Together, these components form a problem-to-module correspondence: UFE extracts robust coral semantics, SM reduces water-medium-induced appearance bias, and IA suppresses illumination-induced feature-statistical bias. This unified design enables reliable segmentation across entirely unseen waters and capture conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI