Skip to content

Latest commit

 

History

History
26 lines (20 loc) · 1.87 KB

HighLevelAlgo.md

File metadata and controls

26 lines (20 loc) · 1.87 KB

High-Level Description of the Algorithm

  1. Data Preparation:

    • Collect and preprocess a dataset of seismic images containing salt domes.
    • Annotate the images with salt body masks, focusing on samples where the salt occupies between 10% and 90% of the total area.
  2. Training the Variational Autoencoder (VAE):

    • Encoder Structure: Build the encoder using a simple feedforward neural network with four stacked dense layers.
    • Latent Space: Encode the seismic image data into a latent space by learning the distribution of key features (e.g., salt boundaries).
    • Decoder Structure: Construct the decoder with five stacked dense layers to decode latent variables into salt body masks.
    • Training Process: Train the VAE using only the annotated salt masks, focusing on learning the boundaries and shapes of salt bodies.
  3. Generating New Salt Body Masks:

    • Latent Sampling: During inference, sample latent variables ( z ) from the learned prior distribution ( \pi(z) ).
    • Mask Generation: Decode the sampled ( z ) using the VAE’s decoder to generate new salt body masks that align with the distribution of the training data.
  4. Contextual Data Augmentation:

    • Use the generated salt body masks as context orientation for augmenting the seismic data.
    • Identify the boundaries between the salt body and surrounding rock. Everything within the boundary is treated as salt.
  5. Texture Synthesis:

    • Apply a non-parametric texture synthesis algorithm to generate new seismic image samples.
    • Ensure the synthesized image preserves the distinct characteristics of seismic images with salt domes, maintaining realistic texture and context zones.
  6. Final Output:

    • Produce new seismic images that combine the generated salt body masks with synthesized textures, enhancing the dataset for further training and analysis.