Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Denoising Low-dose Images Using Deep Learning of Time Series Images

Yang Shao, Toshie Yaguchi, Toshiaki Tanigaki

TL;DR

This work tackles denoising in ultra-low-dose electron microscopy by combining a pre-training/fine-tuning paradigm with synthetic data generation and a novel 3D synthesis technique. The core idea is to decompose time-series images into spatial and temporal components, then denoise along three orthogonal directions (X, Y, T) and fuse results to produce a stable, high-resolution reconstruction. A U-Net-based base model is trained on synthetic noise and further calibrated with real, low-noise high-resolution data, with three directional denoisers trained separately and averaged for final output. Experiments show the approach outperforms standard baselines in mean-squared error and improves continuity and artifact suppression through 3D synthesis, suggesting practical benefits for time-resolved microscopy and related low-dose imaging domains. The method paves the way for more accurate, robust reconstructions in materials science, biology, and medical diagnostics under challenging illumination conditions.

Abstract

Digital image devices have been widely applied in many fields, including scientific imaging, recognition of individuals, and remote sensing. As the application of these imaging technologies to autonomous driving and measurement, image noise generated when observation cannot be performed with a sufficient dose has become a major problem. Machine learning denoise technology is expected to be the solver of this problem, but there are the following problems. Here we report, artifacts generated by machine learning denoise in ultra-low dose observation using an in-situ observation video of an electron microscope as an example. And as a method to solve this problem, we propose a method to decompose a time series image into a 2D image of the spatial axis and time to perform machine learning denoise. Our method opens new avenues accurate and stable reconstruction of continuous high-resolution images from low-dose imaging in science, industry, and life.

Denoising Low-dose Images Using Deep Learning of Time Series Images

TL;DR

This work tackles denoising in ultra-low-dose electron microscopy by combining a pre-training/fine-tuning paradigm with synthetic data generation and a novel 3D synthesis technique. The core idea is to decompose time-series images into spatial and temporal components, then denoise along three orthogonal directions (X, Y, T) and fuse results to produce a stable, high-resolution reconstruction. A U-Net-based base model is trained on synthetic noise and further calibrated with real, low-noise high-resolution data, with three directional denoisers trained separately and averaged for final output. Experiments show the approach outperforms standard baselines in mean-squared error and improves continuity and artifact suppression through 3D synthesis, suggesting practical benefits for time-resolved microscopy and related low-dose imaging domains. The method paves the way for more accurate, robust reconstructions in materials science, biology, and medical diagnostics under challenging illumination conditions.

Abstract

Digital image devices have been widely applied in many fields, including scientific imaging, recognition of individuals, and remote sensing. As the application of these imaging technologies to autonomous driving and measurement, image noise generated when observation cannot be performed with a sufficient dose has become a major problem. Machine learning denoise technology is expected to be the solver of this problem, but there are the following problems. Here we report, artifacts generated by machine learning denoise in ultra-low dose observation using an in-situ observation video of an electron microscope as an example. And as a method to solve this problem, we propose a method to decompose a time series image into a 2D image of the spatial axis and time to perform machine learning denoise. Our method opens new avenues accurate and stable reconstruction of continuous high-resolution images from low-dose imaging in science, industry, and life.
Paper Structure (12 sections, 2 equations, 4 figures, 1 table)

This paper contains 12 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Research Method
  3. Pre-training fine-tuning paradigm
  4. Three-dimensional (3D) synthesis
  5. Artificial data for base model
  6. Structure of base model
  7. Training of base model
  8. Comparisons
  9. Experiments
  10. Comparative Analysis of Denoising Techniques
  11. Comparative Analysis of 3D synthesis
  12. Conclusion

Figures (4)

  • Figure 1: Schematic diagram of three-dimensional slices
  • Figure 2: Actual captured images and images with artificial noise
  • Figure 3: The particles in the red box maintain the continuity and stability of the display in 4 consecutive frames. (Line 1: Noisy in-put. Line 2: Actual observed image before adding noise. Line 3: Denoising result using only the x-y direction model. Line 4: Denoising results after 3D synthesis)
  • Figure 4: Phantom particles appeared in the first frame of the denoising result using only the x-y direction model. (Line 1: Noisy input. Line 2: Actual observed image before adding noise. Line 3: Denoising result using only the x-y direction model. Line 4: Denoising results after 3D synthesis.)