Planning

Tesseract release planning

Here we can plan the next releases of Tesseract.

Future releases

Here are some ideas for future Tesseract releases.

• Modernize the code (see discussions here and here).

• Replace more Tesseract data types by C++ standard types (GenericVector, ...), especially for the API.

4.0.0

That will be the next release.

See also the discussion for issue #1423.

Open issues which should be fixed

• Issues with the "bug" label (see list here)

• Noise characters recognized with bbox as the entire page #1192

• Segmentation fault when using integer models for LSTM training #1573

• Report a warning when the Tesseract initialisation code detects an unsupported locale setting. (See comment)

• Insufficient error message when output file cannot be created Issue 1424

• “no best words!!” on mixed language (fra+ara) items (see issue 235)

• mgr_.Init(traineddata_path.c_str()):Error:Assert failed: #1075 (see issue 1075)

Features wanted for this release

• Script for installing only selected languages from github (see issue)

  https://github.com/zdenop/tessdata_downloader

To be discussed

Depending on available resources and opinions, these suggestions will either be added to the planning for the next or a future release or abandoned.

• Enhance --list-langs to show additional information for scripts and languages like legacy / LSTM, version
  

  This will make the command slower, because each file must be opened and parsed. Add this as --list-langs-details or as --list-lang-details for one language file based on lang-code?

• --list-langs should also display the directory it is using

• Fix the autotools build so that the debug mode uses -O0 as intended

• Add option to optionally select implementation for dot product (CPU, SSE, AVX, ...)

• Relative includes for traineddata

  tessedit_load_sublangs should search for the sublangs relative to the parent, not starting in tessdata dir.

• More fixes for compiler warnings and issues reported by Coverity Scan

• Add a simple bash script for building tesseract

• New traineddata format

  In addition to the current proprietary format Tesseract could also support ZIP archives (see discussion). A possible implementation using libarchive is available, but needs more testing.

• "Training light" - Learning by doing (see issue)

• Modify text2image to use PrepareDistortedPix() #1052

• Schedule date

Regression of features from 3.0x

Tesseract 4.0 should be a full replacement for Tesseract 3.05 and have the same features when used with the old OCR engine (--oem 0). The following regressions still need verification (are they really regressions, or are they just missing features for LSTM):

• User Words (See comment)

• User Patterns (See issue)

Features from 3.0x which are missing for LSTM

These features still work with the old OCR engine (--oem 0), but are missing and desired for LSTM.

• Black list / White list (See issue). Here is a workaround.

• hOCR font info (See comment)

Future release

Here we collect important issues and features for the release(s) following 4.0.0.

• New LSTM-based OSD detector (see comment).

• Remove Legacy Tesseract Engine (see issue)

• Better Multi-language implementation for training (See comment)

• ARM SIMD support for dot product #519

• Using OpenMP for dot product #983

• Remove deprecated code

  This does not include OpenCL or the old Tesseract engine.

• Tesseract creates output for missing input (see issue 1023).

  Mostly solved, but could be improved.

• Issue 1353: Patch for /training/tessopt.cpp (see pull request 13)

  It looks like it is not possible to run more than one training in the same process. The pull request describes a possible fix, but does not include a complete implementation (low priority).