Project a new optical character recognition system is being developed, supporting over 40 languages including English, German, French, Japanese, Chinese, Korean, Uzbek, Azerbaijani, and Lithuanian. Cyrillic-based languages are not currently supported, but their addition is planned. The code is written in Python using the framework and under the Apache 2.0 license. For download pre-trained models for Latin alphabet-based languages and hieroglyphs.
Machine learning methods are used for text detection and recognition in images. A machine learning algorithm is used for text detection (Character-Region Awareness For Text) in for PyTorch, capable of highlighting text on arbitrary objects including labels, information signs, and road signs. A convolutional recurrent neural network is used for character sequence recognition (Convolutional Recurrent Neural Network, a combination of DCNN and RNN) and the CTC BeamSearch (Connectionist Temporal Classification) for decoding the neural network's output into a text representation.
Source: opennet.ru
