The implemention of NPT, Disentangling Noise Pattern from Seismic Images: Noise Reduction and Style Transfer .
NPT has two key parts, which are named I2I-NT and D2D-NT. I2I-NT provides image to image noise transfer while D2D-NT trains dataset to dataset noise transfer using the same network structure with DnCNN.
For new readers, we recommend they follow the steps below to better understand our model:
-
- Build conda environment using requirments.txt
-
- Read & run i2i_nt.ipynb to see how I2I-NT works
-
- Read d2d_nt.py, then checking its configuration in /options
-
- Run & check comparation.ipynb to see the results
The directory structure and files of this project is detailed as:
Directory | Description |
---|---|
/data | Seismic data processing method for D2D-NT training |
/fault_interpretation | Experiments for transferability |
/model_zoo | Npt models that are trained by us, which outputs results in paper. |
/models | Network structures of D2D-NT and baselines |
/options | Hyper-parameter and configuration of D2D-NT model |
/parameter_test | Data patches that used in our experiments |
/tdtv_patches | Samples that ouputs by TDTV |
/utils | Models that used for seismic image processing |
comparation.ipynb | Experiment results in our paper |
d2d_nt.py | Code of D2D-NT model |
i2i_nt.ipynb | Code of I2I-NT model. We also provide some examples for readers to fine-tune the parameters on their datasets. |
image_utils.py | Utils that used for experiments |
no_clean.png | Position image |
no_noise.png | Position image |
requirements.txt | Readers can use this file to build a conda runtime environment |
ricker.ipynb | Method to generate our FSSynth dataset. |
tdtv.py | TDTV model that implemented by us |
tdtv_validate | The smoothing process implemention and examples of TDTV model |