LDPFNet: a lightweight dual-stage polarization fusion network for underwater image descattering

Underwater polarization imaging provides physical cues for separating target-reflected light from backscattered light, but existing methods either rely on fragile model assumptions or underuse polarization information in deep networks. In this paper, we propose LDPFNet, a lightweight dual-stage polarization fusion network for underwater image descattering. Given the intensity image together with DoLP and AoP, LDPFNet first performs global descattering with a polarization-guided encoder-decoder and then refines high-frequency details at full resolution. The network integrates PSFT for conditional polarization modulation, HCAB for balanced scattering suppression and texture preservation, DSAM for supervised inter-stage feature regulation, and FRDNet for dense detail recovery. Experiments on public and natural underwater datasets demonstrate accurate restoration and strong generalization, achieving 30.9521 dB PSNR and 0.8713 SSIM with only 6.606 M parameters. This work provides an effective solution for underwater visual perception in turbid environments, with potential utility in underwater exploration, inspection, and search-and-rescue applications.