De-noising imaging through diffusers with autocorrelation.

Area of Science:

• Optical Imaging
• Computational Imaging
• Machine Learning for Optics

Background:

• Target recovery in optical imaging is challenging due to noise interference in imperfect environments.
• Existing methods often require high signal-to-noise ratio (SNR) speckle patterns and struggle with de-noising and generalizability.
• Autocorrelation information from high-SNR patterns is a key physical constraint for target recovery.

Purpose of the Study:

• To propose a robust, data-driven, two-stage method for de-noising and reconstructing targets from low-SNR speckle patterns.
• To leverage high-SNR autocorrelation information as a physical constraint within a convolutional neural network (CNN).
• To improve the generalizability of target recovery across various diffusers and noise types.

Main Methods:

• A two-stage convolutional neural network (CNN), termed autocorrelation reconstruction (ACR) CNN, was designed.
• The first stage focuses on de-noising, while the second stage reconstructs the target.
• The method incorporates high-SNR autocorrelation information as a physical constraint to guide the learning process.

Main Results:

• The de-noising stage significantly improved peak SNR from 20 dB to 38 dB in system data.
• The reconstruction stage successfully recovered targets from low-SNR speckle patterns with detector and photon noise, outperforming unconstrained methods.
• Experimental validation demonstrated robustness across various diffusers and noise levels, including simulated and real optical system noise.

Conclusions:

• The proposed two-stage ACR CNN method effectively improves target recovery robustness and generalizability in optical imaging.
• Incorporating physical constraints like autocorrelation optimizes the learning process for de-noising and reconstruction.
• This approach shows significant potential for applications in low-illumination imaging and other challenging optical scenarios.