Construction of universal russian characters classifiers BASED ON CONVOLUTIONAL NEURAL NETWORKs

Abstract

Presents the method of constructing universal classifiers of text images based on three-level committee of convolutional neural networks, generation and accounting of specific text patterns features, depending on methods of their synthesis. Efficiency of this method was confirmed by creation classifiers of Russian language characters, accuracy of which exceeds level of leading commercial counterpart in recognition of representative database patterns created in course of research.

Keywords

• распознавание образов
• сверточная нейронная сеть
• обучение
• комитет
• универсальность
• способ синтеза
• генерация
• база