This website uses cookies to improve your experience. If you continue without changing your settings, you consent to our use of cookies in accordance with our cookie policy. You can disable cookies at any time.



In this work, we propose an approach to tackle the problem of three-dimensional post-stack seismic data compression using multi-resolution deep autoencoders, by training one specific compression networ per seismic volume. We feed a residual convolution neural network with 2D slices extracted from 3D seismic blocks and combine multiple scales to generate a discrete latent representation of the input signal. The transformed latent map has its bits oriented in order to minimize both entropy and bit rate. The decoder performs upsampling followed by convolutions to generate the reconstructed image. With this approach, we achieved compressions at low bit rates, attaining a high-quality reconstruction. Experiments show that our method can effectively compress seismic datasets yielding PSNR values over 40dB at a rate of 33:1.

Presentation Date: Tuesday, October 13, 2020

Session Start Time: 1:50 PM

Presentation Time: 3:30 PM

Location: Poster Station 1

Presentation Type: Poster