DASR-Net: dual-attention scattering restoration network for imaging in turbid media via weakly supervised learning

Optical imaging through turbid media suffers from severe scattering-induced degradation, while fully supervised deep learning methods are constrained by the impracticality of acquiring extensive paired datasets. To address these bottlenecks, we propose DASR-Net, a dual-attention scattering restoration network trained via a novel weakly supervised framework. Our method leverages a physics-informed pseudo-label generation pipeline (ICDP) to overcome the limitation of paired data availability. DASR-Net incorporates frequency-selective and spatial attention modules to effectively mitigate scattering effects in both the frequency and spatial domains. To further validate the robustness of the model, we also constructed a milk-water suspension dataset for generalization testing. Quantitative results indicate that DASR-Net achieves an average PSNR of 22.41 dB and SSIM of 0.744 across varying scattering concentrations, outperforming the leading baseline method with an average PSNR gain of 0.86 dB. Furthermore, the model demonstrates sound cross-domain generalization, achieving a PSNR of 16.63 dB and an SSIM of 0.604 on the unseen milk-water dataset without fine-tuning, confirming that it learns scattering-invariant restoration features.