Imaging through scattering media based on semi-supervised learning.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Imaging through scattering media is challenging due to light diffusion.
• Traditional methods often require precise calibration or paired datasets.
• Developing robust and less-invasive imaging techniques is crucial for various applications.

Purpose of the Study:

• To introduce a novel image-to-image translation method for imaging through scattering media.
• To leverage semi-supervised learning with unlabeled data for enhanced image reconstruction.
• To demonstrate the efficacy of the proposed method in a practical experimental setup.

Main Methods:

• Utilized a cycle generative adversarial network (CycleGAN) for image-to-image translation.
• Employed semi-supervised learning with an unlabeled dataset for training.
• Experimentally validated the method by reconstructing images through diffusers using a spatial light modulator.

Main Results:

• Successfully reconstructed object images that were obscured by scattering media.
• Demonstrated the capability of CycleGAN to learn image mappings without paired training data.
• Achieved less-invasive imaging by bypassing the need for direct, unobstructed views.

Conclusions:

• The proposed CycleGAN-based method offers a powerful approach for imaging through scattering media.
• Semi-supervised learning with unlabeled data is effective for enhancing image reconstruction in challenging optical environments.
• This technique holds potential for advancing less-invasive imaging applications in various scientific and medical fields.