Deringing and denoising in extremely under-sampled Fourier single pixel imaging

Area of Science:

• Optics and Photonics
• Image Processing
• Artificial Intelligence

Background:

• Fourier single pixel imaging (FSI) enables real-time applications through undersampling, but this leads to ringing artifacts (Gibbs phenomenon) and noise, degrading image quality.
• Conventional FSI methods struggle with artifact reduction, especially when using low-grade sensors and reduced measurement strategies.

Purpose of the Study:

• To develop a fast image reconstruction framework for real-time FSI that mitigates artifacts and enhances image quality.
• To leverage deep learning for artifact removal and detail recovery in undersampled FSI data.

Main Methods:

• A deep convolutional autoencoder network (DCAN) was proposed for image reconstruction in FSI.
• The DCAN was trained to learn context from FSI artifacts, enabling deringing and denoising.
• The framework was tested on 256x256 images reconstructed from very low sampling rates (1-4%).

Main Results:

• The proposed DCAN-based FSI framework successfully reduced ringing artifacts and noise.
• The method recovered fine details in the reconstructed images, significantly improving overall image quality.
• Experimental results demonstrated superior performance compared to conventional FSI methods, particularly at low sampling rates.

Conclusions:

• Deep learning, specifically DCAN, offers a powerful solution for enhancing real-time FSI by addressing undersampling artifacts and noise.
• The proposed framework achieves high-quality imaging at extremely low sampling rates, broadening the applicability of FSI in real-time scenarios.