Anti-noise computational imaging using unsupervised deep learning.

Area of Science:

• Computational imaging
• Deep learning
• Image reconstruction

Background:

• Traditional multi-pixel cameras face limitations in challenging imaging scenarios.
• Single-pixel computational imaging systems often suffer from noise and reduced quality under adverse conditions like low photon counts and undersampling.

Purpose of the Study:

• To develop an unsupervised deep learning (UnDL) based anti-noise approach for computational imaging.
• To overcome the limitations of conventional methods in reconstructing images from noisy, undersampled single-pixel measurements.
• To alleviate the difficulty of model training by eliminating the need for clean experimental data.

Main Methods:

• Implementation of an unsupervised deep learning (UnDL) framework for noise suppression in computational imaging.
• Utilizing a novel approach that does not require pre-training with clean experimental data, particularly beneficial for biomedical imaging.
• Validation of the method's performance against conventional single-pixel computational imaging techniques.

Main Results:

• The UnDL approach significantly outperforms traditional methods in reconstructing images corrupted by noise.
• The trained model demonstrates generalization capabilities, successfully imaging a real biological sample.
• High-speed imaging (20 fps) of 64x64 resolution objects was achieved at a low sampling ratio (5%).

Conclusions:

• The proposed UnDL method offers a robust solution for noise suppression in computational imaging.
• This technique effectively enhances image quality in challenging environments, especially for biomedical applications.
• The generalizable nature of the model supports its application in various computational imaging solvers for high-quality results.