U-Net: Convolutional Networks for Biomedical Image Segmentation the original author's code is here
our changes:
1,the conv's padding is same,the paper is valid
2,the loss is the original binary_crossentropy,the paper is weighted binary_crossentropy
- Tensorflow>1.2
pip install tensorflow-gpu
- Keras >= 1.0
pip install keras
- libtiff,can be download here,also we add it in libtiff-0.4.2.tar.gz
python2 setup.py install
the data is ISBI datasets,can be downloaded here
also we have added the dataset in the folder ISBI,there are three .tif
python2 split_merge_tif.py
then you will get images in data folder(30 train images and labels,30 test images)
python3 train_unet.py
finally,you will get train acc(%96.8+),val acc(%91+),val loss(0.2)
python3 test_unet.py
https://zhuanlan.zhihu.com/p/26659914