PAC-Net: physics-consistent and color-aware maritime image dehazing via adaptive haze modeling and progressive learning

Maritime image dehazing is a critical prerequisite for ocean engineering and uncrewed surface vessels (USVs). However, it faces unique and complex challenges compared to terrestrial dehazing, primarily due to spatially inhomogeneous haze distribution, severe blue-green color casts caused by water spectral absorption, and significant cross-domain diversity across varying marine environments. Conventional dehazing models, which are typically trained on uniformly synthesized datasets and rely on magnitude-based attention mechanisms, struggle to restore structural details and color fidelity in such complex scenarios. To address these fundamental limitations, this paper proposes a novel, comprehensive maritime image dehazing framework. First, we introduce a physics-driven Adaptive HazeData (AHD) generation mechanism that incorporates global depth priors and local spatial modulation to synthesize highly realistic, non-uniform maritime haze. Second, we develop a Color-Aware Channel Attention (CCA) module that explicitly integrates global color statistical features to perceive and dynamically correct persistent marine color imbalances. Finally, to handle high cross-domain diversity, we employ a scenario-progressive continuous learning (CL) strategy, allowing the network to sequentially absorb multi-scene priors without suffering from feature averaging or catastrophic forgetting. Extensive experiments are conducted on our proposed Adaptive Maritime Haze Dataset (AMHD) and the standard RESIDE benchmark. The quantitative and qualitative results demonstrate that our proposed framework achieves state-of-the-art performance, significantly outperforming existing methods in structural preservation and color restoration while exhibiting robust cross-domain generalization.