Unsupervised cross talk suppression for self-interference digital holography.

Area of Science:

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

Background:

• Self-interference digital holography enables non-coherent imaging but suffers from crosstalk, limiting resolution.
• Deep learning offers nonlinear modeling for crosstalk suppression but requires extensive paired datasets, which are difficult to obtain experimentally.

Purpose of the Study:

• To develop an unsupervised method for crosstalk suppression in self-interference digital holography.
• To improve the resolution of reconstructed images in low-coherence and partially coherent imaging applications.

Main Methods:

• Proposed an unsupervised crosstalk suppression method using a Cycle-Consistent Generative Adversarial Network (CycleGAN).
• Introduced a saliency constraint to the CycleGAN model, named crosstalk suppressing with unsupervised neural network (CS-UNN).
• Enabled learning of image domain mapping without paired training data, preventing image content distortion.

Main Results:

• Successfully suppressed crosstalk information in reconstructed images from self-interference digital holography.
• Demonstrated effective performance without the need for large paired datasets or complex training strategies.
• Validated the method's capability to avoid distortions in image content.

Conclusions:

• The unsupervised CS-UNN method provides an effective solution for crosstalk suppression in self-interference digital holography.
• This approach overcomes the limitations of paired data acquisition for deep learning in real-world holographic imaging.
• Paves the way for wider application of self-interference digital holography in fields like fluorescence and scattered light imaging.