Polarization-constrained Global-local gated adaptive fusion network for underwater polarization imaging

To address the challenge of simultaneously restoring global low-frequency information and local high-frequency details in underwater polarization image restoration tasks, this paper proposes a polarization-constrained global-local gated adaptive fusion network. The network adaptively adjusts the fusion weights between global background intensity information and local detail features through a gating mechanism, thereby enhancing the overall underwater image restoration performance. A polarization attention mechanism, constructed based on polarization angle (AoP) information, is introduced to strengthen the model's focus on detail-rich regions. Additionally, a polarization loss function is designed to optimize overall image quality while preserving high-frequency polarization characteristics. We establish an unpolarized illumination underwater polarization image dataset for validation. Experimental results demonstrate that the proposed network effectively improves image sharpness, contrast, and structural integrity, outperforming existing state-of-the-art methods. Notably, it exhibits robust restoration capabilities across varying turbidity levels—low, medium, and high—on both high and low reflectance targets.