Compressive Image Recovery
Seismic image acquisition can be costly and time consuming. We adopted an appropriately designed Wasserstein generative adversarial network on compressed seismic image recovery. We first trained a pixel inpainting network on several historical surveys, and then propose a non-uniform sampling recommendation based on the evaluation of reconstructed seismic images and metrics. Our results demonstrated a runtime approximately 300 times faster than the conventional method, and better seismic reconstruction accuracy than the original GAN network.
Keywords: Compressive Sensing, Seismic, Non-uniform Sampling, WGANs, Inpainting
- Summer Intern Project at Anadarko
- Mentors: Dr.Nikolaos Mitsakos, Dr.Ping Lu
- Poster: [NIPS 2019 Workshop]
- Paper: [The leading Edge Jornal]
- Code: Python (Pytorch)
Li, X.R., Mitsakos, N., Lu, P., Xiao, Y., Zhan, C. and Zhao, X., Generative Inpainting Network Applications on Seismic Image Compression and Non-Uniform Sampling. Workshop on Neural Information Processing Systems (NIPS): Solving Inverse Problems with Deep Networks (2019).
Li, X. R., Mitsakos, N., Lu, P., Xiao, Y., & Zhao, X. (2019). Seismic compressive sensing by generative inpainting network: Toward an optimized acquisition survey. The Leading Edge, 38(12), 923-933.