ZF_UNET_224 Pretrained Model

Modification of convolutional neural net "UNET" for image segmentation in Keras framework

Requirements

Python 3.*, Keras 2.1, Tensorflow 1.4

Usage

from zf_unet_224_model import ZF_UNET_224, dice_coef_loss, dice_coef
from keras.optimizers import Adam

model = ZF_UNET_224(weights='generator')
optim = Adam()
model.compile(optimizer=optim, loss=dice_coef_loss, metrics=[dice_coef])

model.fit(...)

Notes

"ZF_UNET_224" Model based on UNET code from following paper: https://arxiv.org/abs/1505.04597
This model used to get 2nd place in DSTL competition: https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection
For training used DICE coefficient: https://en.wikipedia.org/wiki/S%C3%B8rensen%E2%80%93Dice_coefficient
Input shape for model is 224x224 (the same as for other popular CNNs like VGG or ResNet)
It has 3 input channels (to process standard RGB (BGR) images). You can change it with variable "INPUT_CHANNELS"
In most cases model ZF_UNET_224 is ok to be used without pretrained weights.
This code should work fine on both Theano and Tensorflow backends. Code prepared for Keras 2.1, if you need code for Keras 1.2 then use this link:

Pretrained weights

Download: Weights for Tensorflow backend ~123 MB (Keras 2.1, Dice coef: 0.998)

Weights were obtained with random image generator (generator code available here: train_infinite_generator.py). See example of images from generator below.

Example of images from generator

Dice coefficient for pretrained weights: ~0.998. See history of learning below:

Log of dice coefficient during training process