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

Area of Science:

• Optics and Photonics
• Computer Vision
• Biomedical Imaging

Background:

• Optical imaging in turbid media is hindered by scattering, degrading image quality.
• Fully supervised deep learning requires extensive paired datasets, which are difficult to obtain for scattering restoration.

Purpose of the Study:

• To develop a robust deep learning model for scattering restoration in optical imaging.
• To overcome the limitations of paired data acquisition in training scattering restoration networks.

Main Methods:

• Proposed DASR-Net, a dual-attention scattering restoration network.
• Implemented a novel weakly supervised framework with a physics-informed pseudo-label generation pipeline (ICDP).
• Incorporated frequency-selective and spatial attention modules for scattering mitigation.

Main Results:

• DASR-Net achieved an average PSNR of 22.41 dB and SSIM of 0.744 across varying scattering concentrations.
• Outperformed leading baseline methods by an average PSNR gain of 0.86 dB.
• Demonstrated cross-domain generalization on an unseen milk-water dataset (PSNR: 16.63 dB, SSIM: 0.604) without fine-tuning.

Conclusions:

• The proposed weakly supervised framework effectively addresses the challenge of limited paired data.
• DASR-Net successfully mitigates scattering effects in both frequency and spatial domains.
• The model learns scattering-invariant features, enabling robust cross-domain generalization for optical imaging restoration.