A denoising framework for 3D and 2D imaging techniques based on photon detection statistics.

Area of Science:

• Computer Vision
• Image Processing
• Photon Counting Imaging

Background:

• Photon Counting Imaging (PCI) enables 3D object data capture in extremely low light.
• Reconstructed 3D images from PCI often suffer degradation due to limited photon detection.
• Improved image restoration is crucial for accurate 3D object recognition in PCI.

Purpose of the Study:

• To propose a novel unsupervised deep learning network for denoising photon-counted 3D sectional images.
• To enhance the quality of 3D reconstructions from low-light imaging conditions.
• To improve object recognition capabilities in challenging imaging scenarios.

Main Methods:

• Development of a fully unsupervised U-Net based deep learning network.
• Integration of skip blocks within the U-Net architecture to extract salient features.
• Symmetric connection of encoder and decoder blocks with skip connections.

Main Results:

• The proposed deep learning network effectively denoises photon-counted 3D images.
• The denoising performance was quantitatively evaluated using peak signal-to-noise ratio (PSNR).
• The deep learning approach demonstrated superior denoising performance compared to the classical Total Variation (TV) denoising algorithm.

Conclusions:

• Unsupervised deep learning offers a powerful solution for restoring degraded 3D images from Photon Counting Imaging.
• The proposed U-Net architecture with skip blocks significantly enhances image quality and aids object recognition.
• This work advances the application of deep learning in low-light 3D imaging and reconstruction.