Fraktur model for OCRopus

This is a character model for recognizing Fraktur font with OCRopus.

With test data that has not been used in the training process it yields an excellent error rate of 1.089%. The 'standard' Fraktur model shows an error rate of 3.14% for the same test data.

In addition to the pyrnn.gz model to be used with ocropus-rpred I have also trained a character model for CLSTM.

Sources

The ground truth images for this model were selected from these historic books:

• Topographische Chronik von Breslau (1805)
• Menzel: Der praktische Maurer (1847) urn:nbn:de:kobv:b4-25126-9
• Kiel city directory (1888)
• Köln city directory (1891)
• Gartenlaube Heft 1, S. 8 (1897)
• Munzinger: Die Japaner (1898) urn:nbn:de:kobv:b4-200905194219
• Frapan: Bittersüß (1891) urn:nbn:de:kobv:b4-200905191440
• Schiller und Oberösterreich (1905) urn:nbn:at:AT-OOeLB-1099695
• Kreis-Kalender für den Kreis Plön (1909) urn:nbn:de:gbv:8:2-2533517
• Das astronomische Weltbild im Wandel der Zeit (1912) urn:nbn:at:AT-OOeLB-4113427
• Egger: Die christliche Mutter (1914) urn:nbn:de:kobv:b4-30011-9
• Was sollen wir kochen? (1915) urn:nbn:at:AT-OOeLB-1184253
• Der Große Krieg in Einzeldarstellungen - Die Winterschlacht in Masuren (1918) urn:nbn:at:AT-OOeLB-1691328
• Preuschen: Yoshiwara (1920) urn:nbn:de:kobv:b4-30090-0
• Ludendorff: Kriegserinnerungen (1921)
• Ehrenbuch der Gefallenen Stuttgarts (1925)
• Zedlitz: Die Zedlitze und ihre Heimat (1925)

For some rare characters (Q, Y, Ä, Ö and Ü) I have generated some synthetic training data with OCRopus-linegen using the Walbaum Fraktur font and words from a German dictionary.

In addition, samples from these documents are included for testing the robustness of the model:

• Peters: Wisch und Umgegend (1898)
• Internationaler Funkentelegraphenvertrag (1906)
• Das Inselschiff, eine Zeitschrift für Freunde der Literatur und des schönen Buches, page 23 (1914)

Special characters

I and J

A problem with Fraktur are the upper-case letters I and J. For many Fraktur fonts they are indistinguishable. For example, compare the I in Im and the J in Jäger:

However, for a neuronal network two indistinguishable but different characters are a big problem. Therefore, it might be better to decide on one letter to avoid the ambiguity. I suggest the uppercase I. As a consequence every J will be recognized as I. The result will be some strange words such as "Iuly".

Using a language model you can easily replace the wrong characters. In German an uppercase J is usually followed by a lowercase vowel.

s, tz, ch, ck and others

Fraktur has a few special characters that usually do not occur in 'normal' text, such as the long-S or the tz ligature. This model (and the ground truth) recognizes the correct Unicode character LATIN SMALL LETTER LONG S. The tz ligature is interpreted as two characters.

Quotation mark

Often you will find two different kinds of quotation marks, the opening lower quotation mark and the closing upper quotation mark.

I decided to distinguish between the two kinds of quotation marks.

• opening: „
• closing: “

Therefore, the image above becomes „Bildhauer“.

Training

ocropy

To train the model from scratch you can use this command:

ocropus-rtrain -c codec.txt -F 2000 -o fraktur "training/*.bin.png"

(Make sure that the OCRopus directory is included in PATH.) If you want to improve the existing model you can start with it:

ocropus-rtrain -c codec.txt -F 2000 --load fraktur.pyrnn.gz -o fraktur "training/*.bin.png"

You can see the training progress in this image:

If you are interested in the amount of ground-truth data and the character accuracy for each book used for the model traning, take a look at the overview-* diagrams in the images folder.