This folder is a refactored / engineering-friendly version of the original notebook-based repo
Deep-Denoising-Autoencoder-for-Seismic-Random-Noise-Attenuation.
What you get:
- Stage-1 supervised training on synthetic noisy/clean pairs (MSE)
- Stage-2 unsupervised/self-supervised training on field data (correlation-based denoising loss)
- Robust correlation loss (epsilon + clipping) + optional MSE anchor
- Trace-wise windowing (patch extraction) + overlap-add reconstruction
- CLI scripts + YAML configs (VSCode-friendly)
See docs/使用说明书.md for a VSCode-first walkthrough.