Deep learning binary semantic segmentation with U-Net and LinkNet

U-Net

Here I compare a modified U-Net (which uses ResNet residual modules) with a modified U-Net that used dilated convolutions. The U-Net+dilated convolution blocks network won the third place in the Kaggle Carvana image segmentation competition.
The Jupyter Notebook and associated model files are in the modified_unet directory.

LinkNet

This is my implementation of LinkNet in Keras. I use this network on the Carvana Image Masking challenge dataset.
The Jupyter Notebook and associated model files are in the linknet directory.

Libraries

These notebooks were developed with the following libraries/versions:

• Python: 3.6.3
• Keras: 2.1.2
• OpenCV