Deep learning approach for denoising low-SNR correlation plenoptic images

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Vision

Background:

• Correlation Plenoptic Imaging (CPI) offers enhanced resolution and depth of field for 3D imaging.
• Conventional CPI is limited by slow acquisition speeds due to the need for high signal-to-noise ratio (SNR).

Purpose of the Study:

• To address the speed limitations of CPI by implementing a Deep Learning approach.
• To improve image quality in CPI using undersampled frame statistics.

Main Methods:

• A Convolutional Neural Network (CNN) model, utilizing a U-Net architecture with a VGG-19 encoder, was trained using transfer learning.
• The model was fed experimental CPI images reconstructed at various sampling ratios.

Main Results:

• The AI model achieved high image quality, with Structural Similarity (SSIM) index values close to 1.
• Performance surpassed traditional denoising methods, especially for low SNR images.
• Acquisition speed was increased by a factor of 20, enabling up to 200 volumetric images per second.

Conclusions:

• This study demonstrates the first successful application of Artificial Intelligence in CPI.
• The AI-driven approach significantly accelerates CPI, paving the way for real-time, scanning-free volumetric imaging.
• Potential applications include neuronal activity monitoring, machine vision, and security systems.