SyntheticLineGenerator

A synthetic Linegenerator for OCR applications. Based on Belval's TextRecognitionDataGenerator (https://github.com/Belval/TextRecognitionDataGenerator) and NVlabs' ocrodeg (https://github.com/NVlabs/ocrodeg).

Clone this repository and use pip install -r requirements.txt

Tested on Windows 10 and Ubuntu 16.04, requires Python 3.x.

How does it work?

In TextRecognitionDataGenerator, run python run.py -c 1000 -i <path>/text/all_strings_and_web.txt to get 1000 randomly generated images with a font of your choice. You can set the font in TextRecognitionDataGenerator/fonts, where historic fonts should be placed in the historic-folder. If more than one font resides in the historic-folder, the generator will switch between the given fonts randomly. Currently, the generator only uses characters and ligatures that were present in the original 1557-dataset (1557-true_character_occurence.ttf)

In FontForge, you will find the .sfd (FontForge projectfile) and .ttf-files for a historic font generated from the 1557-Methodus-Clenardus dataset. Note that 1557-artifically_enhanced_all_chars contains characters that were not present in the original dataset but have been "composed" of others, e.g. W is composed from 2x V. Note that as of now, only TrueType-Fonts are supported.

In text, you will find .txt-files with all words from the dataset. All words from the 1557-dataset can be found in TextRecognitionGenerator/dicts/hist.txt. For best performance, all_strings_and_web.txt should be used as input, as it contains randomly shuffled lines of length 5 that consist of words from the 1557 dataset and are enriched with latin text from the internet.

The most important parameters

The default parameters will make the LineGenerator output all the generated lines in a format with height and margin that fit Calamari (https://github.com/Calamari-OCR/calamari) and 5 words per line. The files will be written to /out. For OCR applications, the most important parameters will be listed below. However, there are many more, as you will see when running python run.py -h.

• -b specify the background. Defaults to white. Might be a feature for future extension (e.g. with old vocal pages).
• -c specify the amount of images that are to be generated. Defaults to 1000.
• -f specify the format (==height, if text is horizontal) of the generated lines. Defaults to 65 px.
• -e specify the extension for the produced images. Defaults to .png.
• -i specify the inputfile. If none is used, words from the hist-dict will be used.
• -m specify the margins for the text with respect to the border. The format is (upper, left, lower, right). Defaults to a format that is well suited for the 1557-dataset.
• -w specify the word-count of the generated lines. Defaults to 5 words per line.
• -z toggle for the creation of a zip-file at the end, for easier handling and upload of the generated lines.
• -tc specify the textcolor. Defaults to #000000 black.
• -sw specify the spacing between words. Defaults to 0.5.
• -sf toggle for the show-font prompt to see the current font in matplotlib. Only supported for historic fonts.
• -ro toggle for rename-output: When set, the output-files will be given unique hex-filenames instead of incremental filenames. Useful when data from several runs will be merged later.
• -rm toggle for deleting old files in the /out-folder before generating new ones. Use with care.

Some minor augmentation features:

• -rk toggle for random skewing of the images, using an angle in the interval [-x,+x], where x is specified with -k
• -rbl toggle for a random blurring with intensity in the interval [-x,+x], where x is specified with -bl

Data Augmentation

The script augment_images.py will apply image augmentation to the given input. The following augmentations will be used:

• random blobs
• rotation and rescaling
• gaußian blurring
• (random noise-) distorsion

Use -f to control the intensity of the augmentation in a range from ]0,10.0]. Note that higher factors make the rescaling-process slow and somewhat useless, as the image will be cropped at the border. This can be resolved by using the -r toggle, that allows to exclude the rotation from the augmentation.

All augmented images will be written to a folder specified by -o (if none is given, /augmentations will be used) and can be zipped by using -z. Use the -s toggle if you want the augmented images to be written to their respective separate folders. To use more than one augmentation run per file, simply run augment_images.py again with the former output as input.

Potential future TO-DOs:

• make character spacing variable (not trivial)
• use background and textcolor from real data instead of plain white and black
• use handwritten model for more variance in generated data