Optical data augmentation enhances out-of-distribution robustness of scattering medium image transmission

Area of Science:

• Computational Imaging
• Deep Learning
• Optical Physics

Background:

• Deep learning excels at reconstructing light fields from scattering media.
• Current methods struggle with out-of-distribution (OOD) data due to high complexity.
• OOD robustness typically requires extensive experimental data collection.

Purpose of the Study:

• To develop a data augmentation strategy for enhanced generalization in scattering media imaging.
• To overcome limitations of current deep learning reconstruction methods under data scarcity.
• To improve the robustness of computational imaging models for unseen scenarios.

Main Methods:

• Proposed an optical image mixing (OIM) approach for optical-domain data augmentation.
• Physically mixed optical images to expand the training data distribution.
• Validated OIM on a multimode fiber (MMF) platform for image reconstruction.

Main Results:

• Achieved generalized reconstruction of 4096-pixel grayscale images with only 200 training samples.
• OIM provided a 40-fold data augmentation, significantly enhancing generalization.
• Improved image reconstruction fidelity by 26.2% compared to conventional models.

Conclusions:

• Optical image mixing (OIM) effectively enhances deep learning model generalization for scattering media imaging.
• OIM serves as a plug-and-play module for improving existing computational imaging reconstruction networks.
• The method significantly reduces the need for extensive experimental data collection.