eynollah/docs/train.md

Training documentation

This document aims to assist users in preparing training datasets, training models, and performing inference with trained models. We cover various use cases including pixel-wise segmentation, image classification, image enhancement, and machine-based reading order detection. For each use case, we provide guidance on how to generate the corresponding training dataset.

The following three tasks can all be accomplished using the code in the train directory:

• Generate training dataset
• Train a model
• Inference with the trained model

Training, evaluation and output

The train and evaluation folders should contain subfolders of images and labels.

The output folder should be an empty folder where the output model will be written to.

Generate training dataset

The script generate_gt_for_training.py is used for generating training datasets. As the results of the following command demonstrates, the dataset generator provides several subcommands:

eynollah-training generate-gt --help

The three most important subcommands are:

• image-enhancement
• machine-based-reading-order
• pagexml2label

image-enhancement

Generating a training dataset for image enhancement is quite straightforward. All that is needed is a set of high-resolution images. The training dataset can then be generated using the following command:

eynollah-training image-enhancement \
  -dis "dir of high resolution images" \
  -dois "dir where degraded images will be written" \
  -dols "dir where the corresponding high resolution image will be written as label" \
  -scs "degrading scales json file"

The scales JSON file is a dictionary with a key named scales and values representing scales smaller than 1. Images are downscaled based on these scales and then upscaled again to their original size. This process causes the images to lose resolution at different scales. The degraded images are used as input images, and the original high-resolution images serve as labels. The enhancement model can be trained with this generated dataset. The scales JSON file looks like this:

{
    "scales": [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9]
}

machine-based-reading-order

For machine-based reading order, we aim to determine the reading priority between two sets of text regions. The model's input is a three-channel image: the first and last channels contain information about each of the two text regions, while the middle channel encodes prominent layout elements necessary for reading order, such as separators and headers. To generate the training dataset, our script requires a PAGE XML file that specifies the image layout with the correct reading order.

For output images, it is necessary to specify the width and height. Additionally, a minimum text region size can be set to filter out regions smaller than this minimum size. This minimum size is defined as the ratio of the text region area to the image area, with a default value of zero. To run the dataset generator, use the following command:

eynollah-training generate-gt machine-based-reading-order \
  -dx "dir of GT xml files" \
  -domi "dir where output images will be written" \
"" -docl "dir where the labels will be written" \
  -ih "height" \
  -iw "width" \
  -min "min area ratio"

pagexml2label

pagexml2label is designed to generate labels from PAGE XML GT files for various pixel-wise segmentation use cases, including:

• printspace (i.e. page frame),
• layout (i.e. regions),
• textline,
• word, and
• glyph.

To train a pixel-wise segmentation model, we require images along with their corresponding labels. Our training script expects a PNG image where each pixel corresponds to a label, represented by an integer. The background is always labeled as zero, while other elements are assigned different integers. For instance, if we have ground truth data with four elements including the background, the classes would be labeled as 0, 1, 2, and 3 respectively.

In binary segmentation scenarios such as textline or page extraction, the background is encoded as 0, and the desired element is automatically encoded as 1 in the PNG label.

To specify the desired use case and the elements to be extracted in the PNG labels, a custom JSON file can be passed. For example, in the case of textline detection, the JSON contents could be this:

{
"use_case": "textline"
}

In the case of layout segmentation, the config JSON file might look like this:

{
"use_case": "layout",
"textregions": {"rest_as_paragraph": 1, "drop-capital": 1, "header": 2, "heading": 2, "marginalia": 3},
"imageregion": 4,
"separatorregion": 5,
"graphicregions": {"rest_as_decoration": 6, "stamp": 7}
}

The same example if PrintSpace (or Border) should be represented as a unique class:

{
"use_case": "layout",
"textregions": {"rest_as_paragraph": 1, "drop-capital": 1, "header": 2, "heading": 2, "marginalia": 3},
"imageregion": 4,
"separatorregion": 5,
"graphicregions": {"rest_as_decoration": 6, "stamp": 7}
"printspace_as_class_in_layout": 8
}

In the layout use-case, it is beneficial to first understand the structure of the PAGE XML file and its elements. For a given page image, the visible segments are annotated in XML with their polygon coordinates and types. On the region level, available segment types include TextRegion, SeparatorRegion, ImageRegion, GraphicRegion, NoiseRegion and TableRegion.

Moreover, text regions and graphic regions in particular are subdivided via @type:

• The allowed subtypes for text regions are paragraph, heading, marginalia, drop-capital, header, footnote, footnote-continued, signature-mark, page-number and catch-word.
• The known subtypes for graphic regions are handwritten-annotation, decoration, stamp and signature.

These types and subtypes must be mapped to classes for the segmentation model. However, sometimes these fine-grained distinctions are not useful or the existing annotations are not very usable (too scarce or too unreliable). In that case, instead of these subtypes with a specific mapping, they can be pooled together by using the two special types:

• rest_as_paragraph (mapping missing TextRegion subtypes and paragraph)
• rest_as_decoration (mapping missing GraphicRegion subtypes and decoration)

(That way, users can extract all known types from the labels and be confident that no subtypes are overlooked.)

In the custom JSON example shown above, header and heading are extracted as the same class, while marginalia is modelled as a different class. All other text region types, including drop-capital, are grouped into the same class. For graphic regions, stamp has its own class, while all other types are classified together. ImageRegion and SeparatorRegion will also represented with a class label in the training data. However, other regions like NoiseRegion or TableRegion will not be included in the PNG files, even if they were present in the PAGE XML.

The tool expects various command-line options:

eynollah-training generate-gt pagexml2label \
  -dx "dir of input PAGE XML files" \
  -do "dir of output label PNG files" \
  -cfg "custom config JSON file" \
  -to "output type (2d or 3d)"

As output type, use

• 2d for training,
• 3d to just visualise the labels.

We have also defined an artificial class that can be added to (rendered around) the boundary of text region types or text lines in order to make separation of neighbouring segments more reliable. The key is called artificial_class_on_boundary, and it takes a list of text region types to be applied to.

Our example JSON config file could then look like this:

{
    "use_case": "layout",
    "textregions": {
        "paragraph": 1,
        "drop-capital": 1,
        "header": 2,
        "heading": 2,
        "marginalia": 3
    },
    "imageregion": 4,
    "separatorregion": 5,
    "graphicregions": {
        "rest_as_decoration": 6
    },
    "artificial_class_on_boundary": ["paragraph", "header", "heading", "marginalia"],
    "artificial_class_label": 7
}

This implies that the artificial class label (denoted by 7) will be present in the generated PNG files and will only be added around segments labeled paragraph, header, heading or marginalia. (This class will be handled specially during decoding at inference, and not show up in final results.)

For printspace, textline, word, and glyph segmentation use-cases, there is no artificial_class_on_boundary key, but artificial_class_label is available. If specified in the config file, then its value should be set at 2, because these elements represent binary classification problems (with background represented as 0, and segments as 1, respectively).

For example, the JSON config for textline detection could look as follows:

{
    "use_case": "textline",
    "artificial_class_label": 2
}

If the coordinates of PrintSpace (or Border) are present in the PAGE XML ground truth files, and one wishes to crop images to only cover the print space bounding box, this can be achieved by passing the -ps option. Note that in this scenario, the directory of the original images must also be provided, to ensure that the images are cropped in sync with the labels. The command line would then resemble this:

eynollah-training generate-gt pagexml2label \
  -dx "dir of input PAGE XML files" \
  -do "dir of output label PNG files" \
  -cfg "custom config JSON file" \
  -to "output type (2d or 3d)" \
  -ps \
  -di "dir of input original images" \
  -doi "dir of output cropped images"

Train a model

classification

For the image classification use-case, we have not provided a ground truth generator, as it is unnecessary. All we require is a training directory with subdirectories, each containing images of its respective classes. We need separate directories for training and evaluation, and the class names (subdirectories) must be consistent across both directories. Additionally, the class names should be specified in the config JSON file, as shown in the following example. If, for instance, we aim to classify "apple" and "orange," with a total of 2 classes, the classification_classes_name key in the config file should appear as follows:

{
    "backbone_type" : "nontransformer",
    "task": "classification",
    "n_classes" : 2,
    "n_epochs" : 10,
    "input_height" : 448,
    "input_width" : 448,
    "weight_decay" : 1e-6,
    "n_batch" : 4,
    "learning_rate": 1e-4,
    "f1_threshold_classification": 0.8,
    "pretraining" : true,
    "classification_classes_name" : {"0":"apple",  "1":"orange"},
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

Then dir_train should be like this:

.
└── train             # train directory
   ├── apple          # directory of images for apple class
   └── orange         # directory of images for orange class

And dir_eval analogously:

.
└── eval              # evaluation directory
   ├── apple          # directory of images for apple class
   └── orange         # directory of images for orange class

The classification model can be trained using the following command line:

eynollah-training train with config_classification.json

As evident in the example JSON file above, for classification, we utilize a "f1_threshold_classification" parameter. This parameter is employed to gather all models with an evaluation f1 score surpassing this threshold. Subsequently, an ensemble of these model weights is executed, and a model is saved in the output directory as "model_ens_avg". Additionally, the weight of the best model based on the evaluation f1 score is saved as "model_best".

reading order

An example config json file for machine based reading order should be like this:

{
    "backbone_type" : "nontransformer",
    "task": "reading_order",
    "n_classes" : 1,
    "n_epochs" : 5,
    "input_height" : 672,
    "input_width" : 448,
    "weight_decay" : 1e-6,
    "n_batch" : 4,
    "learning_rate": 1e-4,
    "pretraining" : true,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

The "dir_train" should be like this:

.
└── train             # train directory
   ├── images          # directory of images
   └── labels         # directory of labels

And the "dir_eval" the same structure as train directory:

.
└── eval             # evaluation directory
   ├── images          # directory of images
   └── labels         # directory of labels

The reading-order model can be trained like the classification case command line.

Segmentation (Textline, Binarization, Page extraction and layout) and enhancement

Parameter configuration for segmentation or enhancement usecases

The following parameter configuration can be applied to all segmentation use cases and enhancements. The augmentation, its sub-parameters, and continued training are defined only for segmentation use cases and enhancements, not for classification and machine-based reading order, as you can see in their example config files.

• backbone_type: For segmentation tasks (such as text line, binarization, and layout detection) and enhancement, we offer two backbone options: a "nontransformer" and a "transformer" backbone. For the "transformer" backbone, we first apply a CNN followed by a transformer. In contrast, the "nontransformer" backbone utilizes only a CNN ResNet-50.
• task: The task parameter can have values such as "segmentation", "enhancement", "classification", and "reading_order".
• 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. In the case of reading_order it should set to 1. And for the case of layout detection just the unique number of classes should be given.
• 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.
• pretraining: Set to true to load pretrained weights of ResNet50 encoder. The downloaded weights should be saved in a folder named "pretrained_model" in the same directory of "train.py" script.
• augmentation: If you want to apply any kind of augmentation this parameter should first set to true.
• flip_aug: If true, different types of filp will be applied on image. Type of flips is given with "flip_index" parameter.
• blur_aug: If true, different types of blurring will be applied on image. Type of blurrings is given with "blur_k" parameter.
• scaling: If true, scaling will be applied on image. Scale of scaling is given with "scales" parameter.
• degrading: If true, degrading will be applied to the image. The amount of degrading is defined with "degrade_scales" parameter.
• brightening: If true, brightening will be applied to the image. The amount of brightening is defined with "brightness" parameter.
• rotation_not_90: If true, rotation (not 90 degree) will be applied on image. Rotation angles are given with "thetha" parameter.
• 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.
• flip_index: Type of flips.
• blur_k: Type of blurrings.
• scales: Scales of scaling.
• brightness: The amount of brightenings.
• thetha: Rotation angles.
• degrade_scales: The amount of degradings.
• 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 providethe dir of trained model with "dir_of_start_model" and index for naming themodels. For example if you have already trained for 3 epochs then your lastindex 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 truethe 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 augmentthem and then wewrite 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.
• index_start: Starting index for saved models in the case that "continue_training" is true.
• dir_of_start_model: Directory containing pretrained model to continue training the model in the case that "continue_training" is true.
• transformer_num_patches_xy: Number of patches for vision transformer in x and y direction respectively.
• transformer_patchsize_x: Patch size of vision transformer patches in x direction.
• transformer_patchsize_y: Patch size of vision transformer patches in y direction.
• transformer_projection_dim: Transformer projection dimension. Default value is 64.
• transformer_mlp_head_units: Transformer Multilayer Perceptron (MLP) head units. Default value is [128, 64].
• transformer_layers: transformer layers. Default value is 8.
• transformer_num_heads: Transformer number of heads. Default value is 4.
• transformer_cnn_first: We have two types of vision transformers. In one type, a CNN is applied first, followed by a transformer. In the other type, this order is reversed. If transformer_cnn_first is true, it means the CNN will be applied before the transformer. Default value is true.

In case of segmentation and enhancement the train and evaluation data should be organised as follows.

The "dir_train" directory should be like this:

.
└── train             # train directory
   ├── images          # directory of images
   └── labels         # directory of labels

And the "dir_eval" the same structure as train directory:

.
└── eval             # evaluation directory
   ├── images          # directory of images
   └── labels         # directory of labels

After configuring the JSON file for segmentation or enhancement, training can be initiated by running the following command line, similar to classification and reading-order model training:

eynollah-training train with config_classification.json

Binarization

Ground truth format

Lables for each pixel are identified by a number. So if you have a binary case, n_classes should be set to 2 and labels should be 0 and 1 for each class and pixel.

In the case of multiclass, just set n_classes to the number of classes you have and the try to produce the labels by pixels set from 0 , 1 ,2 .., n_classes-1. The labels format should be png. Our lables are 3 channel png images but only information of first channel is used. If you have an image label with height and width of 10, for a binary case the first channel should look like this:

Label: [ [1, 0, 0, 1, 1, 0, 0, 1, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         ...,
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0] ] 

This means that you have an image by 10*10*3 and pixel[0,0] belongs to class 1 and pixel[0,1] belongs to class 0.

A small sample of training data for binarization experiment can be found here, Training data sample, which contains images and lables folders.

An example config json file for binarization can be like this:

{
    "backbone_type" : "transformer",
    "task": "binarization",
    "n_classes" : 2,
    "n_epochs" : 4,
    "input_height" : 224,
    "input_width" : 672,
    "weight_decay" : 1e-6,
    "n_batch" : 1,
    "learning_rate": 1e-4,
    "patches" : true,
    "pretraining" : true,
    "augmentation" : true,
    "flip_aug" : false,
    "blur_aug" : false,
    "scaling" : true,
    "degrading": false,
    "brightening": false,
    "binarization" : false,
    "scaling_bluring" : false,
    "scaling_binarization" : false,
    "scaling_flip" : false,
    "rotation": false,
    "rotation_not_90": false,
    "transformer_num_patches_xy": [7, 7],
    "transformer_patchsize_x": 3,
    "transformer_patchsize_y": 1,
    "transformer_projection_dim": 192,
    "transformer_mlp_head_units": [128, 64],
    "transformer_layers": 8,
    "transformer_num_heads": 4,
    "transformer_cnn_first": true,
    "blur_k" : ["blur","guass","median"],
    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
    "brightness" : [1.3, 1.5, 1.7, 2],
    "degrade_scales" : [0.2, 0.4],
    "flip_index" : [0, 1, -1],
    "thetha" : [10, -10],
    "continue_training": false,
    "index_start" : 0,
    "dir_of_start_model" : " ",
    "weighted_loss": false,
    "is_loss_soft_dice": false,
    "data_is_provided": false,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

Textline

{
    "backbone_type" : "nontransformer",
    "task": "segmentation",
    "n_classes" : 2,
    "n_epochs" : 4,
    "input_height" : 448,
    "input_width" : 224,
    "weight_decay" : 1e-6,
    "n_batch" : 1,
    "learning_rate": 1e-4,
    "patches" : true,
    "pretraining" : true,
    "augmentation" : true,
    "flip_aug" : false,
    "blur_aug" : false,
    "scaling" : true,
    "degrading": false,
    "brightening": false,
    "binarization" : false,
    "scaling_bluring" : false,
    "scaling_binarization" : false,
    "scaling_flip" : false,
    "rotation": false,
    "rotation_not_90": false,
    "blur_k" : ["blur","guass","median"],
    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
    "brightness" : [1.3, 1.5, 1.7, 2],
    "degrade_scales" : [0.2, 0.4],
    "flip_index" : [0, 1, -1],
    "thetha" : [10, -10],
    "continue_training": false,
    "index_start" : 0,
    "dir_of_start_model" : " ",
    "weighted_loss": false,
    "is_loss_soft_dice": false,
    "data_is_provided": false,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

Enhancement

{
    "backbone_type" : "nontransformer",
    "task": "enhancement",
    "n_classes" : 3,
    "n_epochs" : 4,
    "input_height" : 448,
    "input_width" : 224,
    "weight_decay" : 1e-6,
    "n_batch" : 4,
    "learning_rate": 1e-4,
    "patches" : true,
    "pretraining" : true,
    "augmentation" : true,
    "flip_aug" : false,
    "blur_aug" : false,
    "scaling" : true,
    "degrading": false,
    "brightening": false,
    "binarization" : false,
    "scaling_bluring" : false,
    "scaling_binarization" : false,
    "scaling_flip" : false,
    "rotation": false,
    "rotation_not_90": false,
    "blur_k" : ["blur","guass","median"],
    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
    "brightness" : [1.3, 1.5, 1.7, 2],
    "degrade_scales" : [0.2, 0.4],
    "flip_index" : [0, 1, -1],
    "thetha" : [10, -10],
    "continue_training": false,
    "index_start" : 0,
    "dir_of_start_model" : " ",
    "weighted_loss": false,
    "is_loss_soft_dice": false,
    "data_is_provided": false,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

It's important to mention that the value of n_classes for enhancement should be 3, as the model's output is a 3-channel image.

Page extraction

{
    "backbone_type" : "nontransformer",
    "task": "segmentation",
    "n_classes" : 2,
    "n_epochs" : 4,
    "input_height" : 448,
    "input_width" : 224,
    "weight_decay" : 1e-6,
    "n_batch" : 1,
    "learning_rate": 1e-4,
    "patches" : false,
    "pretraining" : true,
    "augmentation" : false,
    "flip_aug" : false,
    "blur_aug" : false,
    "scaling" : true,
    "degrading": false,
    "brightening": false,
    "binarization" : false,
    "scaling_bluring" : false,
    "scaling_binarization" : false,
    "scaling_flip" : false,
    "rotation": false,
    "rotation_not_90": false,
    "blur_k" : ["blur","guass","median"],
    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
    "brightness" : [1.3, 1.5, 1.7, 2],
    "degrade_scales" : [0.2, 0.4],
    "flip_index" : [0, 1, -1],
    "thetha" : [10, -10],
    "continue_training": false,
    "index_start" : 0,
    "dir_of_start_model" : " ",
    "weighted_loss": false,
    "is_loss_soft_dice": false,
    "data_is_provided": false,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

For page segmentation (or print space or border segmentation), the model needs to view the input image in its entirety,hence the patches parameter should be set to false.

layout segmentation

An example config json file for layout segmentation with 5 classes (including background) can be like this:

{
    "backbone_type" : "transformer",
    "task": "segmentation",
    "n_classes" : 5,
    "n_epochs" : 4,
    "input_height" : 448,
    "input_width" : 224,
    "weight_decay" : 1e-6,
    "n_batch" : 1,
    "learning_rate": 1e-4,
    "patches" : true,
    "pretraining" : true,
    "augmentation" : true,
    "flip_aug" : false,
    "blur_aug" : false,
    "scaling" : true,
    "degrading": false,
    "brightening": false,
    "binarization" : false,
    "scaling_bluring" : false,
    "scaling_binarization" : false,
    "scaling_flip" : false,
    "rotation": false,
    "rotation_not_90": false,
    "transformer_num_patches_xy": [7, 14],
    "transformer_patchsize_x": 1,
    "transformer_patchsize_y": 1,
    "transformer_projection_dim": 64,
    "transformer_mlp_head_units": [128, 64],
    "transformer_layers": 8,
    "transformer_num_heads": 4,
    "transformer_cnn_first": true,
    "blur_k" : ["blur","guass","median"],
    "scales" : [0.6, 0.7, 0.8, 0.9, 1.1, 1.2, 1.4],
    "brightness" : [1.3, 1.5, 1.7, 2],
    "degrade_scales" : [0.2, 0.4],
    "flip_index" : [0, 1, -1],
    "thetha" : [10, -10],
    "continue_training": false,
    "index_start" : 0,
    "dir_of_start_model" : " ",
    "weighted_loss": false,
    "is_loss_soft_dice": false,
    "data_is_provided": false,
    "dir_train": "./train",
    "dir_eval": "./eval",
    "dir_output": "./output"
}

Inference with the trained model

classification

For conducting inference with a trained model, you simply need to execute the following command line, specifying the directory of the model and the image on which to perform inference:

eynollah-training inference -m "model dir" -i "image"

This will straightforwardly return the class of the image.

machine based reading order

To infer the reading order using a reading order model, we need a PAGE XML file containing layout information but without the reading order. We simply need to provide the model directory, the XML file, and the output directory. The new XML file with the added reading order will be written to the output directory with the same name. We need to run:

eynollah-training inference \
  -m "model dir" \
  -xml "page xml file" \
  -o "output dir to write new xml with reading order"

Segmentation (Textline, Binarization, Page extraction and layout) and enhancement

For conducting inference with a trained model for segmentation and enhancement you need to run the following command line:

eynollah-training inference \
  -m "model dir" \
  -i "image" \
  -p \
  -s "output image"

Note that in the case of page extraction the -p flag is not needed.

For segmentation or binarization tasks, if a ground truth (GT) label is available, the IoU evaluation metric can be calculated for the output. To do this, you need to provide the GT label using the argument -gt.