Spectrum Enhancement (SE) targets reconstruction and denoising of noisy spectral or microscopy observations. In materials characterization workflows, low-dose or low-signal acquisitions often reduce image quality and make atomic-scale structure analysis harder. SE models learn to recover cleaner signals from paired noisy and reference data, improving downstream inspection of crystal structures, defects, and local material morphology.
The current PaddleMaterials SE workflow focuses on STEM image enhancement.
Given noisy HAADF or BF STEM inputs, the model predicts a configured target
image (gt_enhance or gt_detect) and supports training, evaluation, and
prediction with the common PaddleMaterials trainer/predictor style.
| Supported Functions | SFIN |
|---|---|
| Support Data Types | |
| STEM images | ✅ |
| HAADF / BF inputs | ✅ |
| Spectrum Enhancement | |
| Image denoising/enhancement | ✅ |
| Detection target restoration | ✅ |
| ML Capabilities · Training | |
| Single-GPU | ✅ |
| Distributed Train | - |
| Mixed Precision | - |
| Fine-tuning | ✅ |
| ML Capabilities · Predict | |
| Standard inference | ✅ |
| Distributed inference | - |
| Dataset | |
| HAADF STEM | ✅ |
| BF STEM | ✅ |
Legend: ✅ Verified · 🧪 Implemented, pending validation · 🚧 In development · - Not supported · 🌟 Original Work