U-Net with ResNet and ResNext Encoder Pre-Trained on ImageNet with HyperColumns, ChannelAttention, SpatialAttentionGate for Image Segmentation

By Mikhail Karchevskiy, Insaf Ashrapov and Leonid Kozinkin

Based on TernausNet.

Introduction

One of the most important applications of seismic reflection is the hydrocarbon exploration which is closely related to salt deposits analysis. This problem is very important even nowadays due to it’s non-linear nature. Taking into account the recent developments in deep learning networks TGS-NOPEC Geophysical Company hosted the Kaggle competition for salt deposits segmentation problem in seismic image data. In this paper, we demonstrate the great performance of several novel deep learning techniques merged into a single neural network. Using a U-Net with ResNeXt-50 encoder pretrained on ImageNet as our base architecture, we implemented Spatial-Channel Squeeze & Excitation, Lovasz loss, analog of CoordConv and Hypercolumn methods.

This architecture was a part of the solutiuon (27th out of 3234 teams top 1%) in the TGS Salt Identification Challenge.

Model architecture

Citing

Please cite in your publications if it helps your research:

@ARTICLE{arXiv:1812.01429,
         author = {Mikhail Karchevskiy, Insaf Ashrapov, Leonid Kozinkin},
          title = {Automatic salt deposits segmentation: A deep learning approach},
        journal = {ArXiv e-prints},
         eprint = {1812.01429},
           year = 2018
        }