Style transfer for generation of realistically textured subsurface models. Visualized and explained.
Ovcharenko Oleg, Vladimir Kazei, Daniel Peter, and Tariq Alkhalifah. "Style transfer for generation of realistically textured subsurface models." In SEG Technical Program Expanded Abstracts 2019, pp. 2393-2397. Society of Exploration Geophysicists, 2019.
The notebook in this repository reproduces the workflow for texture-transfer from an elastic isotropic subsurface model to a prior synthetic distribution. We follow the (Gatys et al., 2015) to transfer texture from a Marmousi II benchmark geological model to a background distribution generated using a random Gaussian field.
Make a random Gaussian field resamble the Marmousi II layered features.
We apply the iterative optimization approach which benefits from higher control at cost of longer generation times. To accelerate the texture transfer one would use a GAN-based approach as proposed by (Johnson et al., 2016) and (Ulyanov et al., 2016).
Seeing the outputs from intermediate layers in the network leads to better understaing of what is going on at each step of the algorithm.
Enforcing the optimization to match certain areas from the style model ultimately leads to a controlled texture generation.
In the terminal, go to the folder with the notebook and run the command
jupyter notebook geo_style.ipynb
https://github.com/kevinzakka/style-transfer
https://github.com/rrmina/neural-style-pytorch
https://www.tensorflow.org/beta/tutorials/generative/style_transfer
Jupyter 4.4.0
Keras 2.2.4
Numpy 1.15.4
Pillow 5.3.0
Scipy 1.1.0
@incollection{ovcharenko2019style,
title={Style transfer for generation of realistically textured subsurface models},
author={Ovcharenko, Oleg and Kazei, Vladimir and Peter, Daniel and Alkhalifah, Tariq},
booktitle={SEG Technical Program Expanded Abstracts 2019},
pages={2393--2397},
year={2019},
publisher={Society of Exploration Geophysicists}
}