The architecture was inspired by Road Extraction by Deep Residual U-Net
Data should be prepared in the PASCAL VOC annotation way. For more information see docsting of PASCALVOCIterator class (utils.py)
For preprocessing you could use standart keras utils for image preprocessing (keras.preprocessing.image)
The architecture of the model based on the Road Extraction by Deep Residual U-Net. Image below illustrates it.
In order to train model you could use train.py file. First of all you need to specify input_shape, dataset_folder and classes variables and then run train.py
This implementation depends on following libraries:
- Tensorflow
- Keras == 2.1.2 (probably >= 2.1.2)