RSS
Логотип
Баннер в шапке 1
Баннер в шапке 2

Smart Engines GreenOCR

Product
The name of the base system (platform): Artificial intelligence (AI, Artificial intelligence, AI)
Developers: Smart Engines (Smart Endzhins)
Last Release Date: 2025/09//16
Technology: EDMS - Streaming Recognition Systems

Main article: OCR - Optical Character Recognition

2025: GreenOCR 2.0

Smart Engines on September 16, 2025 introduced GreenOCR 2.0, the green AI generation for text recognition. The ultralight computing architectures used in it - 4.6-bit and averaging 8-bit neural networks - made it possible to read data faster by 20%, as well as reduce power consumption during recognition. At the same time, the number of errors compared to the previous version decreased by 10 times.

GreenOCR is based on the technology of replacing real numbers with integers - this allows you to deploy all the functionality of AI even on devices without GPUs. The second generation (version 2.0) combines two technologies - 4.6-bit and averaging 8-bit networks created specifically for modern mobile devices of the ARM architecture. Filters of 4,6-bit neural networks are quantized in such a way that intermediate data has limited bit, allowing more efficient use of vector operations of the processor. And in the computing core of averaging 8-bit neural networks, a dot product based on tree averaging is used, which allows you to use registers twice as intense. The simultaneous use of these two ideas made it possible to achieve unrivaled recognition speed.

Improvements in speed and energy efficiency were made not to the detriment of quality. Moreover, this model is mistaken 10 times less when recognizing documents relative to the previous version. These indicators were provided by neural networks with an expanded alphabetical space, trained to understand not only the symbols themselves, but also their structural relationships: the proximity of letter fragments, the points of connection of elements and other graphic features. The system works confidently with printed and handwritten text in 103 languages.

File:Aquote1.png
GreenOCR 2.0 meets the key Green AI criteria adopted in the scientific literature: low power consumption, minimal carbon dioxide emissions, and green learning approaches. Due to the completely local operation of the technology, data centers and megawatts of energy are not needed: all recognition of the company's customers emit less CO2 per year than 3 outdoor gas heaters for the same period, - said Vladimir Arlazarov, CEO of Smart Engines, Doctor of Technical Sciences.
File:Aquote2.png