gganssle / cseg-imxlate

CSEG Image Translation

Welcome to the repo for Graham Ganssle's article, Denoising Seismic by Image Translation Networks, in the CSEG Recorder, January 2018 edition! If you have any questions, please send Graham a message.

Results

It works! The above image is an example output from the denoising network.

The above image is a comparison of the denoising network with three common denoising filters: total variation, bilateral, and wavelet.

Setup

To run this code you'll need Torch7 running on CUDA, and CuDNN. You'll also need Agile's bruges, which can be installed simply with pip install bruges.

Running this code

I've included the prep_images.sh script to prep the data for you. To run it use bash prep_images.sh. It'll create a folder, dat/pairs, which has the subdirectories test, train, and val. These three folders are the ones you should use to train your model.

Training will take a considerable amount of compute. To do a full training run with 200 epochs it took a whole day on an NVIDIA Tesla K80.

Acknowledgements

The image translation network used in this project was the wonderful pix2pix. Thanks Phillipi!

Thanks, also, to Steve Purves for building the Earth models provided (in .zip format) in dat/zips! To use these models simply cd dat/zips/ and unzip xxx, where xxx is the file name.