4.0 with LSTM

4.0

Tesseract 4.0 alpha source code is available in the 'master' branch of the repository. It adds a new OCR engine based on LSTM neural networks. It initially works (well) on x86/Linux. Model data for 101 languages is available in the tessdata repository.

Documentation

• NeuralNetsInTesseract4.00

• VGSLSpecs

• DAS 2016 tutorial slides
  Slides #2, #6, #7 have information about LSTM integration in Tesseract 4.0.

• 4.0 Accuracy and Performance

Training Tesseract LSTM engine

• TrainingTesseract 4.00
• TrainingTesseract 4.00 - Finetuning Example - Arabic
• TrainingTesseract 4.00 - Replace Top Layer Example - Norwegian
• TrainingTesseract 4.00 - Replace Top Layer Example - Devanagari

3.0 version of box files can be converted for use with LSTM training by adding a tab character at end of each line. Mark EOL function under Edit in Box Editor tab of latest version of JTessboxeditor can be used to do it automatically.

4.0.0-alpha ppa

Unofficial Ubuntu PPAs for Tesseract 4.00 & Leptonica 1.74:

• https://launchpad.net/~alex-p/+archive/ubuntu/tesseract-ocr

Leponica 1.74.1 package for Debian:

• https://packages.debian.org/sid/libleptonica-dev

4.0.0-alpha for Windows

Unofficial experimental binaries of tesseract-ocr 4.0.0-alpha (Jan 30, 2017) are available from the following links:

• Windows Installer made with MinGW-w64 from UB Mannheim
• zip file with cppan generated .dll and .exe files, You have to install VC2015 x86 redist from microsoft.com in order to run them.

4.0.0-alpha with GUI frontend

VietOCR

Windows binaries of tesseract-ocr 4.0.0-alpha with GUI interface are available for VietOCR from

• VietOCR5.0alpha

• Visual C++ Redistributable for Visual Studio 2015 runtime - vc_redist.x86.exe is REQUIRED for VietOCR to run correctly.

VietOCR can be used to download appropriate 4.0.0alpha traineddata for additional languages.

gImageReader

Windows binaries of tesseract-ocr 4.0.0-alpha with GUI interface are available for gImageReader from

• gImageReader_3.2.1_qt5_i686_tesseract4.0.0.git2f10be5.exe
• gImageReader_3.2.1_qt5_x86_64_tesseract4.0.0.git2f10be5.exe

Download 4.0.0alpha traineddata to use with the above from master branch of tessdata. e.g. for Hindi download the following file:

https://github.com/tesseract-ocr/tessdata/blob/master/hin.traineddata *

3.05-dev

An unofficial installer for Tesseract 3.05-dev for Windows is available from [Tesseract at UB Mannheim] (https://github.com/UB-Mannheim/tesseract/wiki). This includes the training tools.

The [3.05 branch on GitHub] (https://github.com/tesseract-ocr/tesseract/tree/3.05) can be used by those who want the bug fixes for 3.04 release.

3.04.1

The current official release is [3.04.1] (https://github.com/tesseract-ocr/tesseract/releases/tag/3.04.01).