Adaptive imaging through dense dynamic scattering media using transfer learning.

Area of Science:

• Optical Imaging
• Deep Learning
• Scattering Media

Background:

• Imaging through scattering media is crucial for biology, transportation, and remote sensing.
• Deep learning methods offer rapid imaging but struggle with generalization, especially in dynamic scattering environments.
• Existing methods lack adaptability to changing scattering conditions.

Purpose of the Study:

• To develop an adaptive imaging method for dense dynamic scattering media.
• To overcome the generalization limitations of conventional deep learning approaches.
• To enable clear imaging under varying and non-static scattering conditions.

Main Methods:

• Utilized transfer learning by fine-tuning a pre-trained de-scattering model.
• Employed a known segment of the imaging target for model adaptation.
• Acquired downstream task training data simultaneously with test data.

Main Results:

• Achieved clear imaging under varying scattering conditions.
• Demonstrated over 5dB performance improvement compared to conventional deep learning.
• Showcased adaptability to optical thickness variations from 11.6 to 13.1.

Conclusions:

• Transfer learning enables adaptive imaging through dense dynamic scattering media.
• The proposed method enhances imaging performance and robustness.
• Holds promise for applications like video surveillance and beacon guidance in challenging environments.