Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Denoising of Two-Phase Optically Sectioned Structured Illumination Reconstructions Using Encoder-Decoder Networks

Allison Davis, Yezhi Shen, Xiaoyu Ji, Fengqing Zhu

TL;DR

This work tackles denoising in two-phase optical-sectioning structured illumination (OS-SI) by leveraging synthetic training data to overcome the lack of clean ground-truth OS-SI images. It compares two encoder-decoder architectures—the asymmetrical denoising autoencoder (DAE) and a Denoising U-Net—trained on artifact-corrupted synthetic images and evaluated on real pool-boiling OS-SI data. Results show both networks improve reconstruction quality, with the DAE excelling at removing structured line artifacts in 2P reconstructions and the U-Net preserving edge details and mitigating nonuniform illumination in SHT reconstructions. The study demonstrates that synthetic data enables supervised denoising of OS-SI and highlights the potential of encoder-decoder networks to streamline reconstruction workflows, while noting the need for broader datasets and domain adaptation for real-world variability.

Abstract

Structured illumination (SI) enhances image resolution and contrast by projecting patterned light onto a sample. In two-phase optical-sectioning SI (OS-SI), reduced acquisition time introduces residual artifacts that conventional denoising struggles to suppress. Deep learning offers an alternative to traditional methods; however, supervised training is limited by the lack of clean, optically sectioned ground-truth data. We investigate encoder-decoder networks for artifact reduction in two-phase OS-SI, using synthetic training pairs formed by applying real artifact fields to synthetic images. An asymmetrical denoising autoencoder (DAE) and a U-Net are trained on the synthetic data, then evaluated on real OS-SI images. Both networks improve image clarity, with each excelling against different artifact types. These results demonstrate that synthetic training enables supervised denoising of OS-SI images and highlight the potential of encoder-decoder networks to streamline reconstruction workflows.

Denoising of Two-Phase Optically Sectioned Structured Illumination Reconstructions Using Encoder-Decoder Networks

TL;DR

This work tackles denoising in two-phase optical-sectioning structured illumination (OS-SI) by leveraging synthetic training data to overcome the lack of clean ground-truth OS-SI images. It compares two encoder-decoder architectures—the asymmetrical denoising autoencoder (DAE) and a Denoising U-Net—trained on artifact-corrupted synthetic images and evaluated on real pool-boiling OS-SI data. Results show both networks improve reconstruction quality, with the DAE excelling at removing structured line artifacts in 2P reconstructions and the U-Net preserving edge details and mitigating nonuniform illumination in SHT reconstructions. The study demonstrates that synthetic data enables supervised denoising of OS-SI and highlights the potential of encoder-decoder networks to streamline reconstruction workflows, while noting the need for broader datasets and domain adaptation for real-world variability.

Abstract

Structured illumination (SI) enhances image resolution and contrast by projecting patterned light onto a sample. In two-phase optical-sectioning SI (OS-SI), reduced acquisition time introduces residual artifacts that conventional denoising struggles to suppress. Deep learning offers an alternative to traditional methods; however, supervised training is limited by the lack of clean, optically sectioned ground-truth data. We investigate encoder-decoder networks for artifact reduction in two-phase OS-SI, using synthetic training pairs formed by applying real artifact fields to synthetic images. An asymmetrical denoising autoencoder (DAE) and a U-Net are trained on the synthetic data, then evaluated on real OS-SI images. Both networks improve image clarity, with each excelling against different artifact types. These results demonstrate that synthetic training enables supervised denoising of OS-SI images and highlight the potential of encoder-decoder networks to streamline reconstruction workflows.

Paper Structure

This paper contains 12 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Description of Structured Illumination Reconstruction
  3. Method
  4. Model Architectures
  5. Asymmetrical Denoising Autoencoder
  6. Denoising U-Net
  7. Datasets
  8. Training Data: Synthetic Bubbles Dataset
  9. Test Data: Structured-Illumination Pool-Boiling Dataset
  10. Noising of Training Data
  11. Experiments
  12. Conclusion

Figures (4)

  • Figure 1: Comparison of denoising techniques on two reconstruction methods. (a) Consecutive images used in reconstruction; (b) 2P (top) and SHT (bottom) reconstructions. (c)-(f) Denoised outputs from notch filtering, BM3D, DnCNN (supervised on synthetic data), and N2N (self-supervised on real OS-SI images). (g) Results from our DAE encoder-decoder network.
  • Figure 2: Network architectures used for artifact reduction. (a) Asymmetrical DAE, consisting of a downsampling encoder, a latent representation, and an upsampling decoder. (b) U-Net architecture with symmetric skip connections, which better preserves spatial detail during reconstruction.
  • Figure 3: Synthetic-bubble images with two-phase OS-SI reconstruction artifacts. (a) Residuals from 2P reconstruction. (b) Residuals derived from SHT reconstruction.
  • Figure 4: Qualitative denoising on two samples of real two-phase OS-SI pool-boiling images. The rows show 2P and SHT reconstructions (top), denoised reconstructions using the asymmetrical DAE (middle), and denoised reconstructions using the U-Net (bottom).