> Perform document layout analysis (segmentation) from image data and return the results as [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML).
This tool performs document layout analysis (segmentation) from image data and returns the results as [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML).
`pip install .` or
`pip install . -e` for editable installation
Alternatively, you can also use `make` with these targets:
`make install` or
`make install-dev` for editable installation
### Models
In order to run this tool you need trained models. You can download our pretrained models from [qurator-data.de](https://qurator-data.de/eynollah/).
Alternatively, running `make models` will download and extract models to `$(PWD)/models_eynollah`.
### Training
In case you want to train your own model to use with Eynollah, have a look at [sbb_pixelwise_segmentation](https://github.com/qurator-spk/sbb_pixelwise_segmentation).
## Usage
The command-line interface can be called like this:
In addition, the tool can be used to detect the _[ReadingOrder](https://ocr-d.de/en/gt-guidelines/trans/lyLeserichtung.html)_ of regions. The final goal is to feed the output to an OCR model.
In addition, the tool can detect the [ReadingOrder](https://ocr-d.de/en/gt-guidelines/trans/lyLeserichtung.html) of regions. The final goal is to feed the output to an OCR model.
</details>
### Method description
<details>
<summary>click to expand/collapse</summary><br/>
The tool uses a combination of various models and heuristics (see flowchart below for the different stages and how they interact):
Eynollah uses a combination of various models and heuristics (see flowchart below for the different stages and how they interact):
For the purpose of text recognition (OCR) and in order to avoid noise being introduced from texts outside the printspace, one first needs to detect the border of the printed frame. This is done by a binary pixel-wise-segmentation model trained on a dataset of 2,000 documents where about 1,200 of them come from the [dhSegment](https://github.com/dhlab-epfl/dhSegment/) project (you can download the dataset from [here](https://github.com/dhlab-epfl/dhSegment/releases/download/v0.2/pages.zip)) and the remainder having been annotated in SBB. For border detection, the model needs to be fed with the whole image at once rather than separated in patches.
For the purpose of text recognition (OCR) and in order to avoid noise being introduced from texts outside the printspace, one first needs to detect the border of the printed frame. This is done by a binary pixel-wise-segmentation model trained on a dataset of 2,000 documents where about 1,200 of them come from the [dhSegment](https://github.com/dhlab-epfl/dhSegment/) project (you can download the dataset from [here](https://github.com/dhlab-epfl/dhSegment/releases/download/v0.2/pages.zip)) and the remainder having been annotated in SBB. For border detection, the model needs to be fed with the whole image at once rather than separated in patches.
## Layout detection
### Layout detection
As a next step, text regions need to be identified by means of layout detection. Again a pixel-wise segmentation model was trained on 131 labeled images from the SBB digital collections, including some data augmentation. Since the target of this tool are historical documents, we consider as main region types text regions, separators, images, tables and background - each with their own subclasses, e.g. in the case of text regions, subclasses like header/heading, drop capital, main body text etc. While it would be desirable to detect and classify each of these classes in a granular way, there are also limitations due to having a suitably large and balanced training set. Accordingly, the current version of this tool is focussed on the main region types background, text region, image and separator.
As a next step, text regions need to be identified by means of layout detection. Again a pixel-wise segmentation model was trained on 131 labeled images from the SBB digital collections, including some data augmentation. Since the target of this tool are historical documents, we consider as main region types text regions, separators, images, tables and background - each with their own subclasses, e.g. in the case of text regions, subclasses like header/heading, drop capital, main body text etc. While it would be desirable to detect and classify each of these classes in a granular way, there are also limitations due to having a suitably large and balanced training set. Accordingly, the current version of this tool is focussed on the main region types background, text region, image and separator.
## Textline detection
#### Textline detection
In a subsequent step, binary pixel-wise segmentation is used again to classify pixels in a document that constitute textlines. For textline segmentation, a model was initially trained on documents with only one column/block of text and some augmentation with regard to scaling. By fine-tuning the parameters also for multi-column documents, additional training data was produced that resulted in a much more robust textline detection model.
In a subsequent step, binary pixel-wise segmentation is used again to classify pixels in a document that constitute textlines. For textline segmentation, a model was initially trained on documents with only one column/block of text and some augmentation with regard to scaling. By fine-tuning the parameters also for multi-column documents, additional training data was produced that resulted in a much more robust textline detection model.
## Image enhancement
#### Image enhancement
This is an image to image model which input was low quality of an image and label was actually the original image. For this one we did not have any GT, so we decreased the quality of documents in SBB and then feed them into model.
This is an image to image model which input was low quality of an image and label was actually the original image. For this one we did not have any GT, so we decreased the quality of documents in SBB and then feed them into model.
## Scale classification
#### Scale classification
This is simply an image classifier which classifies images based on their scales or better to say based on their number of columns.
This is simply an image classifier which classifies images based on their scales or better to say based on their number of columns.
## Heuristic methods
### Heuristic methods
Some heuristic methods are also employed to further improve the model predictions:
Some heuristic methods are also employed to further improve the model predictions:
* After border detection, the largest contour is determined by a bounding box, and the image cropped to these coordinates.
* After border detection, the largest contour is determined by a bounding box, and the image cropped to these coordinates.
* For text region detection, the image is scaled up to make it easier for the model to detect background space between text regions.
* For text region detection, the image is scaled up to make it easier for the model to detect background space between text regions.
@ -53,50 +114,39 @@ Some heuristic methods are also employed to further improve the model prediction
* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels.
* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels.
* Finally, using the derived coordinates, bounding boxes are determined for each textline.
* Finally, using the derived coordinates, bounding boxes are determined for each textline.
## Installation
</details>
`pip install .` or
`pip install . -e` for editable installation
### Model description
Alternatively, you can also use `make` with these targets:
<details>
<summary>click to expand/collapse</summary><br/>
`make install` or
TODO
`make install-dev` for editable installation
</details>
### Models
### How to use
In order to run this tool you also need trained models. You can download our pretrained models from [qurator-data.de](https://qurator-data.de/eynollah/).
<details>
<summary>click to expand/collapse</summary><br/>
Alternatively, running `make models` will download and extract models to `$(PWD)/models_eynollah`.
First, this model makes use of up to 9 trained models which are responsible for different operations like size detection, column classification, image enhancement, page extraction, main layout detection, full layout detection and textline detection.That does not mean that all 9 models are always required for every document. Based on the document characteristics and parameters specified, different scenarios can be applied.
## Usage
* If none of the parameters is set to `true`, the tool will perform a layout detection of main regions (background, text, images, separators and marginals). An advantage of this tool is that it tries to extract main text regions separately as much as possible.
The basic command-line interface can be called like this:
* If you set `-ae` (**a**llow image **e**nhancement) parameter to `true`, the tool will first check the ppi (pixel-per-inch) of the image and when it is less than 300, the tool will resize it and only then image enhancement will occur. Image enhancement can also take place without this option, but by setting this option to `true`, the layout xml data (e.g. coordinates) will be based on the resized and enhanced image instead of the original image.
```sh
* For some documents, while the quality is good, their scale is very large, and the performance of tool decreases. In such cases you can set `-as` (**a**llow **s**caling) to `true`. With this option enabled, the tool will try to rescale the image and only then the layout detection process will begin.
* If you care about drop capitals (initials) and headings, you can set `-fl` (**f**ull **l**ayout) to `true`. With this setting, the tool can currently distinguish 7 document layout classes/elements.
* In cases where the document includes curved headers or curved lines, rectangular bounding boxes for textlines will not be a great option. In such cases it is strongly recommended setting the flag `-cl` (**c**urved **l**ines) to `true` to find contours of curved lines instead of rectangular bounding boxes. Be advised that enabling this option increases the processing time of the tool.
* To crop and save image regions inside the document, set the parameter `-si` (**s**ave **i**mages) to true and provide a directory path to store the extracted images.
The tool does accept and works better on original images (RGB format) than binarized images.
* This tool is actively being developed. If problems occur, or the performance does not meet your expectations, we welcome your feedback via [issues](https://github.com/qurator-spk/eynollah/issues).
### `--full-layout` vs `--no-full-layout`
#### `--full-layout` vs `--no-full-layout`
Here are the difference in elements detected depending on the `--full-layout`/`--no-full-layout` command line flags:
Here are the difference in elements detected depending on the `--full-layout`/`--no-full-layout` command line flags:
@ -111,20 +161,22 @@ Here are the difference in elements detected depending on the `--full-layout`/`-
| marginals / text line | x | x |
| marginals / text line | x | x |
| image region | x | x |
| image region | x | x |
### How to use
#### Use as OCR-D processor
First, this model makes use of up to 9 trained models which are responsible for different operations like size detection, column classification, image enhancement, page extraction, main layout detection, full layout detection and textline detection.That does not mean that all 9 models are always required for every document. Based on the document characteristics and parameters specified, different scenarios can be applied.
Eynollah ships with a CLI interface to be used as [OCR-D](https://ocr-d.de) processor. In this case, the source image file group with (preferably) RGB images should be used as input like this:
* If none of the parameters is set to `true`, the tool will perform a layout detection of main regions (background, text, images, separators and marginals). An advantage of this tool is that it tries to extract main text regions separately as much as possible.
* If you set `-ae` (**a**llow image **e**nhancement) parameter to `true`, the tool will first check the ppi (pixel-per-inch) of the image and when it is less than 300, the tool will resize it and only then image enhancement will occur. Image enhancement can also take place without this option, but by setting this option to `true`, the layout xml data (e.g. coordinates) will be based on the resized and enhanced image instead of the original image.
In fact, the image referenced by `@imageFilename` in PAGE-XML is passed on directly to Eynollah as a processor, so that e.g. calling
* For some documents, while the quality is good, their scale is very large, and the performance of tool decreases. In such cases you can set `-as` (**a**llow **s**caling) to `true`. With this option enabled, the tool will try to rescale the image and only then the layout detection process will begin.
* If you care about drop capitals (initials) and headings, you can set `-fl` (**f**ull **l**ayout) to `true`. With this setting, the tool can currently distinguish 7 document layout classes/elements.
would still use the original (RGB) image despite any binarization that may have occured in previous OCR-D processing steps
* In cases where the document includes curved headers or curved lines, rectangular bounding boxes for textlines will not be a great option. In such cases it is strongly recommended setting the flag `-cl` (**c**urved **l**ines) to `true` to find contours of curved lines instead of rectangular bounding boxes. Be advised that enabling this option increases the processing time of the tool.
#### Eynollah "light"
* To crop and save image regions inside the document, set the parameter `-si` (**s**ave **i**mages) to true and provide a directory path to store the extracted images.
TODO
* This tool is actively being developed. If problems occur, or the performance does not meet your expectations, we welcome your feedback via [issues](https://github.com/qurator-spk/eynollah/issues).