Pre-processed speckle-correlation learning for imaging through scattering media

Area of Science:

• Optics
• Image Processing
• Machine Learning

Background:

• Imaging through scattering media is crucial for applications like biomedical imaging and remote sensing.
• Deep learning methods for scattering imaging often require high-quality real-world data, limiting their use in resource-limited scenarios.

Purpose of the Study:

• To develop a pre-processed speckle-correlation learning technique to improve scattering imaging.
• To enhance model generalization and reduce dependence on real-world training data.

Main Methods:

• A pre-processed speckle-correlation learning technique was developed.
• The method decreases data distribution differences under various scattering conditions.
• The network acquires reliable physical prior knowledge for enhanced generalization.

Main Results:

• The proposed approach achieved better structural similarity index and peak signal-to-noise ratio compared to conventional methods.
• Demonstrated superior stability against unknown scattering media with different statistical distributions.
• Successfully recovered real-world data using only simulated data for training.

Conclusions:

• The technique offers a robust and feasible solution for scattering imaging in resource-scarce environments.
• Eliminates the need for real-world data, overcoming a major limitation in current deep learning approaches.
• Significantly enhances the generalization ability of scattering imaging models.