Researchers from the Siberian Federal University (SFU) and the St. Petersburg State Electrotechnical University “LETY”, subordinate departments of the Ministry of Education of Russia, have developed a new convolutional neural network (CNN). It can recognize handwritten letters of the Russian alphabet in an image with an accuracy of up to 99%.
The neural network was trained by pre-processing data from the CoMNIST repository containing handwritten Latin and Cyrillic alphabets. The data set for analysis included 13,299 photographs with large, printed, and italic letters. For programming, the scientists chose the Python language and the interactive Jupyter development environment.
As reported on the website of the Ministry of Education and Science, the researchers first built a new data set and labeled images for 33 letters of the Russian alphabet. They then built a CNN architecture to detect these letters and compared the results with existing powerful CNN models. Finally, a complete description of the convolutional neural network and the source code have already been presented for further reproduction by other researchers.
In the future, experts plan to train the neural network to “read” words and sentences as a whole, as well as the ability to distinguish writing styles.