Research map / Image generation / Image restoration
Noise Calibration and Spatial-Frequency Interactive Network for STEM Image Enhancement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Calibrates STEM noise from real atomic images, synthesizes a more realistic and general HAADF/BF STEM dataset, and trains a spatial-frequency interactive network (SFIN) for STEM image enhancement.
Problem
STEM image enhancement suffers from synthetic training data that lacks realistic noise (scan noise depends on image gradients, so flat-field calibration fails), limited atomic arrangement diversity and imaging modes, and enhancement networks that ignore frequency-domain periodicity of atoms.
Approach
Calibrate background, scan, and pointwise noise parameters from real STEM images containing atoms (rather than flat fields), use them to synthesize randomized HAADF and BF datasets, and enhance images with a network that exchanges information between spatial and FFT-based frequency branches.
Key results
- SFIN achieves 38.48 dB PSNR / 0.9582 SSIM on the HAADF synthetic dataset, versus 36.50 dB for FCN, and 31.76 dB on TEMImageNet versus 26.18 dB for FCN.
- The calibrated dataset gives lower KL divergence to real-data noise parameters than GAN and TEMImageNet datasets (e.g., background brightness KLD 1.26 vs 4.66 for TEMImageNet in HAADF).
- Removing frequency-domain processing drops HAADF PSNR from 38.48 to 34.61 dB in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SFIN instead of FCN | STEM image enhancement · TEMImageNet | PSNR: 26.18 -> 31.76 | Helped |
| use SFIN instead of FCN | STEM image enhancement · Our Dataset (HAADF mode) | PSNR: 36.50 -> 38.48 | Helped |
| use SFIN instead of FCN | STEM image enhancement · Our Dataset (BF mode) | PSNR: 32.22 -> 33.24 | Helped |
| add frequency-domain processing (ablation) | SFIN · Our Dataset (HAADF mode) | PSNR: 34.61 -> 38.48 | Helped |
| add frequency-domain processing (ablation) | SFIN · Our Dataset (HAADF mode) | SSIM: 0.9050 -> 0.9582 | Helped |
| add information exchange between spatial and frequency domains (ablation) | SFIN · Our Dataset (HAADF mode) | PSNR: 38.38 -> 38.48 | Helped |
| add information exchange between spatial and frequency domains (ablation) | SFIN · Our Dataset (HAADF mode) | SSIM: 0.9514 -> 0.9582 | Helped |
Surprising findings
- Scan noise standard deviation is proportional to image gradient amplitude, so flat-field calibration methods used for natural images do not transfer to STEM.
Limitations
- Real-data evaluation uses a small set of collected images with known structures and theoretical GT renderings.
- Pure frequency-domain information cannot represent single-atom absence or addition, so spatial branch is required.
- Calibration uses only 40 real images (20 per mode).
Experimental setup
Datasets: TEMImageNet GAN dataset (Cycle-GAN STEM) synthetic STEM dataset (ours) real STEM images (collected, unprocessed)
Benchmarks: TEMImageNet own synthetic STEM dataset (HAADF and BF)
Metrics: PSNR SSIM Kullback-Leibler divergence R2
Baselines: ABSF Wiener filter AtomSegNet FCN GAN dataset TEMImageNet
Builds on: Tukey lambda noise calibration Perlin noise abTEM AtomSegNet
Closest papers by meaning
- A Gaussian Filter-Based 3D Registration Method for Series Section Electron Microscopy (AAAI 2025 · similarity 0.81)
- Learning Large-Factor EM Image Super-Resolution with Generative Priors (CVPR 2024 · similarity 0.81)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.79)
- AMGenC: Generating Charge Balanced Amorphous Materials (NeurIPS 2026 · similarity 0.79)
- LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.78)
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.