This repository contains the source code for training an encoder model for document image segmentation.
## Installation
Either clone the repository via `git clone https://github.com/qurator-spk/sbb_pixelwise_segmentation.git` or download and unpack the [ZIP](https://github.com/qurator-spk/sbb_pixelwise_segmentation/archive/master.zip).
> Tool to extract 2-D or 3-D RGB images from PAGE-XML data. In the former case, the output will be 1 2-D image array which each class has filled with a pixel value. In the case of a 3-D RGB image,
each class will be defined with a RGB value and beside images, a text file of classes will also be produced.
> Extract region classes and their colours in mask (pseg) images. Allows the color map as free dict parameter, and comes with a default that mimics PageViewer's coloring for quick debugging; it also warns when regions do overlap.
A small sample of training data for binarization experiment can be found here, [Training data sample](https://qurator-data.de/~vahid.rezanezhad/binarization_training_data_sample/), which contains images and lables folders.
* patches: If you want to break input images into smaller patches (input size of the model) you need to set this parameter to ``true``. In the case that the model should see the image once, like page extraction, patches should be set to ``false``.
* n_batch: Number of batches at each iteration.
* n_classes: Number of classes. In the case of binary classification this should be 2.
* n_epochs: Number of epochs.
* input_height: This indicates the height of model's input.
* input_width: This indicates the width of model's input.
* weight_decay: Weight decay of l2 regularization of model layers.
* augmentation: If you want to apply any kind of augmentation this parameter should first set to ``true``.
* rotation: If ``true``, 90 degree rotation will be applied on image.
* binarization: If ``true``,Otsu thresholding will be applied to augment the input data with binarized images.
* scaling_bluring: If ``true``, combination of scaling and blurring will be applied on image.
* scaling_binarization: If ``true``, combination of scaling and binarization will be applied on image.
* scaling_flip: If ``true``, combination of scaling and flip will be applied on image.
* continue_training: If ``true``, it means that you have already trained a model and you would like to continue the training. So it is needed to provide the dir of trained model with "dir_of_start_model" and index for naming the models. For example if you have already trained for 3 epochs then your last index is 2 and if you want to continue from model_1.h5, you can set "index_start" to 3 to start naming model with index 3.
* weighted_loss: If ``true``, this means that you want to apply weighted categorical_crossentropy as loss fucntion. Be carefull if you set to ``true``the parameter "is_loss_soft_dice" should be ``false``
* data_is_provided: If you have already provided the input data you can set this to ``true``. Be sure that the train and eval data are in "dir_output". Since when once we provide training data we resize and augment them and then we write them in sub-directories train and eval in "dir_output".
* dir_train: This is the directory of "images" and "labels" (dir_train should include two subdirectories with names of images and labels ) for raw images and labels. Namely they are not prepared (not resized and not augmented) yet for training the model. When we run this tool these raw data will be transformed to suitable size needed for the model and they will be written in "dir_output" in train and eval directories. Each of train and eval include "images" and "labels" sub-directories.
