Repository for converting segmentation images of placenta to realistic looking ones using pix2pix network. There are two main files avilable:
- pix2pix_placenta.ipynb in jupyter notebook format, for training network on the dataset.
- main.py for testing results from network including single image conversion and biger image convesion using stiching.
Model used for converting is avaliable in kaggle here
Original paper can be found here
[Here pic of model]
Model was trained on modified dataset avaliable in kaggle here both in .keras and .h5 formats.
Original data comes form here
Images from dataset are rgb images with black (0,0,0) background and segmented vessels (1,1,1) to better distinguish vessels form background image pixels are multipiled by 40.
generator() model gets as input images of size 256x256 that are cast to tf.float32 data type, than images are normalized and None batch dimension is added to fit to expected model input.