Extrapolated speckle-correlation imaging with an untrained deep neural network.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Speckle-correlation imaging recovers objects through scattering media using the memory effect.
• Current methods are limited by the memory effect's field of view and often assume sparse objects due to ill-posedness.
• Extending the field of view typically involves extrapolating speckle correlation during reconstruction.

Purpose of the Study:

• To overcome the limitations of conventional speckle-correlation imaging for non-sparse objects.
• To extend the field of view in speckle-correlation imaging beyond the memory effect limit.
• To introduce a regularization method that addresses the ill-posed nature of imaging non-sparse objects.

Main Methods:

• Implemented a deep image prior, utilizing an untrained convolutional neural network, to regularize image statistics.
• Integrated the deep image prior into the speckle-correlation imaging reconstruction process.
• Experimentally validated the proposed method for imaging through scattering media.

Main Results:

• Successfully demonstrated speckle-correlation imaging of spatially non-sparse objects.
• Showcased the ability to extend the field of view beyond the inherent memory effect limitations.
• Validated the efficacy of the deep image prior in improving image reconstruction quality.

Conclusions:

• The proposed deep image prior method significantly enhances speckle-correlation imaging capabilities.
• This advancement allows for the observation of complex, non-sparse objects through scattering media.
• The method shows promise for broader applications in imaging through scattering environments.