Adaptive inverse mapping: a model-free semi-supervised learning approach towards robust imaging through dynamic...

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Imaging through scattering media is challenging due to light distortion.
• Dynamic changes in scattering media further complicate image reconstruction.
• Existing methods have limitations regarding prior knowledge and medium accessibility.

Purpose of the Study:

• To develop a novel method for high-quality imaging through dynamic scattering media.
• To overcome limitations of current inverse mapping techniques.
• To demonstrate robustness without prior knowledge of scattering changes.

Main Methods:

• Proposed an adaptive inverse mapping (AIP) method.
• Utilized unsupervised learning to correct inverse mapping using output speckle images.
• Validated the method through numerical simulations (evolving transmission matrix, phase mask) and experimental application (multimode fiber).

Main Results:

• The AIP method successfully reconstructed images through dynamic scattering.
• Achieved increased imaging robustness across all tested scenarios.
• Demonstrated high imaging performance without requiring prior knowledge of dynamic changes.

Conclusions:

• The adaptive inverse mapping (AIP) method offers a robust solution for imaging through dynamic scattering media.
• Unsupervised learning corrects inverse mapping effectively with continuous output monitoring.
• The AIP method shows significant potential for real-world applications in dynamic scattering imaging.