The release of the optical character recognition system Tesseract 5.2 has been published, supporting UTF-8 character recognition and texts in over 100 languages, including Russian, Kazakh, Belarusian, and Ukrainian. The output can be saved as plain text or in HTML (hOCR), ALTO (XML), PDF, and TSV formats. The system was originally created between 1985 and 1995 in the laboratory of Hewlett Packard, with the code released under the Apache license in 2005 and further developed with the involvement of Google employees. The source texts of the project are distributed under the Apache 2.0 license.
Tesseract includes a console utility and the libtesseract library for embedding text recognition features into other applications. Notable third-party GUI interfaces supporting Tesseract include gImageReader, VietOCR, and YAGF. It offers two recognition engines: the classic one, which recognizes text by matching individual character templates, and a new one based on a recurrent neural network (LSTM) machine learning system, optimized for recognizing entire lines of text and significantly improving accuracy. Pre-trained models are available for 123 languages. Performance optimization modules using OpenMP and SIMD instructions such as AVX2, AVX, AVX512F, NEON, or SSE4.1 are provided.
Key improvements in Tesseract 5.2:
- Optimizations have been added using Intel AVX512F instructions.
- A function has been implemented in the C API for initializing Tesseract with loading the machine learning model from memory.
- An invert_threshold parameter has been added, defining the level of inverting text strings. The default value is set to 0.7. To disable inversion, set the value to 0.
- Processing of very large documents on 32-bit hosts has been improved.
- Transition from std::regex functions to std::string has been made.
- Build scripts for Autotools, CMake, and continuous integration systems have been improved.
Source: opennet.ru
