The purpose of this project is to detect salt in seismic images by classifying each pixel in a 101x101 picture. The dataset contains 4,000 training images (with corresponding masks that identify salt) and 18,000 test images. This project was completed as part of the Kaggle TGS Salt competition which closed in October 2018. https://www.kaggle.com/c/tgs-salt-identification-challenge. I teamed up with @des137 for this project.
We used a CNN with U-Net architecture in our approach. A detailed description of the U-Net can be found in the following publication by Olaf Ronneberger, Philipp Fischer and Thomas Brox: https://pdfs.semanticscholar.org/0704/5f87709d0b7b998794e9fa912c0aba912281.pdf. A basic U-Net implementation in Keras is provided by @yihui-he. In our best solution we used number of filters increasing from 32 to 1024 and image sizes decreasing from 128x128 to 4x4 during contracting. We used 4 convolutional layers per U-Net step, filter sizes of 3x3 and 2x2, random normal weight initialization, adam optimizer with learning rate 1e-4, no dropout, and a batch size of 64.
We achieved a mean intersection over union of correctly identified salt of 0.776 on the test set. Below are some examples of predicions on the validation set. Images in the first column are actual seismic images. The second and third columns show the corresponding labeled masks and the predicted masks, respectively. White pixels in the masks represent identified salt.
