Traineddata Files for Version 4.00 +
We have three sets of official .traineddata files trained at Google, for tesseract versions 4.00 and above. These are made available in three separate repositories.
- tessdata_fast (Sep 2017) best “value for money” in speed vs accuracy,
Integermodels. - tessdata_best (Sep 2017) best results on Google’s eval data, slower,
Floatmodels. These are the only models that can be used as base for finetune training. - tessdata (Nov 2016 and Sep 2017) These have legacy tesseract models from 2016. The LSTM models have been updated with Integer version of tessdata_best LSTM models. (Cube based legacy tesseract models for Hindi, Arabic etc. have been deleted).
When using the traineddata files from the tessdata_best and tessdata_fast repositories, only the new LSTM-based OCR engine (–oem 1) is supported. The legacy tesseract engine (–oem 0) is NOT supported with these files, so Tesseract’s oem modes ‘0’ and ‘2’ won’t work with them.
- Special Data Files
- Latest Data Files - Sept. 2017
- Data Files for Version 4.00 - Nov. 2016
- Format of traineddata files
Special Data Files
| Lang Code | Description | 4.x/3.0x traineddata |
|---|---|---|
| osd | Orientation and script detection | osd.traineddata |
| equ | Math / equation detection | equ.traineddata |
Note: These two data files are compatible with older versions of Tesseract. osd is compatible with version 3.01 and up, and equ is compatible with version 3.02 and up.
Updated Data Files (September 15, 2017)
We have three sets of .traineddata files on GitHub in three separate repositories. These are compatible with Tesseract 4.0x+ and 5.0.0.Alpha.
| Trained models | Speed | Accuracy | Supports legacy | Retrainable | |
|---|---|---|---|---|---|
| tessdata | Legacy + LSTM (integerized tessdata-best) | Faster than tessdata-best | Slightly less accurate than tessdata-best | Yes | No |
| tessdata-best | LSTM only (based on langdata) | Slowest | Most accurate | No | Yes |
| tessdata-fast | Integerized LSTM of a smaller network than tessdata-best | Fastest | Least accurate | No | No |
Most users will want tessdata_fast and that is what will be shipped as part of Linux distributions.
tessdata_best is for people willing to trade a lot of speed for slightly better accuracy. It is also the only set of files which can be used for certain retraining scenarios for advanced users.
The third set in tessdata is the only one that supports the legacy recognizer. The 4.00 files from November 2016 have both legacy and older LSTM models. The current set of files in tessdata have the legacy models and newer LSTM models (integer versions of 4.00.00 alpha models in tessdata_best).
Note: When using the new models in the tessdata_best and tessdata_fast repositories, only the new LSTM-based OCR engine is supported. The legacy engine is not supported with these files, so Tesseract’s oem modes ‘0’ and ‘2’ won’t work with them.
Data Files for Version 4.00 (November 29, 2016)
tessdata tagged 4.0.0 has the models from Sept 2017 that have been updated with Integer versions of tessdata_best LSTM models. This set of traineddata files has support for the legacy recognizer with –oem 0 and for LSTM models with –oem 1.
tessdata tagged 4.00 has the models from 2016. The individual language files are linked in the table below.
Note: The kur data file was not updated from 3.04. For Fraktur, use the newer data files from the tessdata_fast or tessdata_best repositories.
Format of traineddata files
The traineddata file for each language is an archive file in a Tesseract specific format. It contains several uncompressed component files which are needed by the Tesseract OCR process. The program combine_tessdata is used to create a tessdata file from the component files and can also extract them again like in the following examples:
Pre 4.0.0 format from Nov 2016 (with both LSTM and Legacy models)
combine_tessdata -u eng.traineddata eng. Extracting tessdata components from eng.traineddata Wrote eng.unicharset Wrote eng.unicharambigs Wrote eng.inttemp Wrote eng.pffmtable Wrote eng.normproto Wrote eng.punc-dawg Wrote eng.word-dawg Wrote eng.number-dawg Wrote eng.freq-dawg Wrote eng.cube-unicharset Wrote eng.cube-word-dawg Wrote eng.shapetable Wrote eng.bigram-dawg Wrote eng.lstm Wrote eng.lstm-punc-dawg Wrote eng.lstm-word-dawg Wrote eng.lstm-number-dawg Wrote eng.version Version string:Pre-4.0.0 1:unicharset:size=7477, offset=192 2:unicharambigs:size=1047, offset=7669 3:inttemp:size=976552, offset=8716 4:pffmtable:size=844, offset=985268 5:normproto:size=13408, offset=986112 6:punc-dawg:size=4322, offset=999520 7:word-dawg:size=1082890, offset=1003842 8:number-dawg:size=6426, offset=2086732 9:freq-dawg:size=1410, offset=2093158 11:cube-unicharset:size=1511, offset=2094568 12:cube-word-dawg:size=1062106, offset=2096079 13:shapetable:size=63346, offset=3158185 14:bigram-dawg:size=16109842, offset=3221531 17:lstm:size=5390718, offset=19331373 18:lstm-punc-dawg:size=4322, offset=24722091 19:lstm-word-dawg:size=7143578, offset=24726413 20:lstm-number-dawg:size=3530, offset=31869991 23:version:size=9, offset=31873521 4.00.00alpha LSTM only format
combine_tessdata -u eng.traineddata eng. Extracting tessdata components from eng.traineddata Wrote eng.lstm Wrote eng.lstm-punc-dawg Wrote eng.lstm-word-dawg Wrote eng.lstm-number-dawg Wrote eng.lstm-unicharset Wrote eng.lstm-recoder Wrote eng.version Version string:4.00.00alpha:eng:synth20170629:[1,36,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx512O1c1] 17:lstm:size=11689099, offset=192 18:lstm-punc-dawg:size=4322, offset=11689291 19:lstm-word-dawg:size=3694794, offset=11693613 20:lstm-number-dawg:size=4738, offset=15388407 21:lstm-unicharset:size=6360, offset=15393145 22:lstm-recoder:size=1012, offset=15399505 23:version:size=80, offset=15400517 Proposal for compressed traineddata files
| There are some proposals to replace the Tesseract archive format by a standard archive format which could also support compression. A [discussion on the tesseract-dev forum](https://groups.google.com/forum/?hl=en#!searchin/tesseract-dev/zip | sort:date/tesseract-dev/U5HSugUeeeI) proposed the ZIP format already in 2014. In 2017 an experimental implementation was provided as a pull request. |
