Generation of Images with Handwritten Text in Russian
The impact of various methods of generating additional training datasets on the quality of recognition models are explored: the method based on handwritten fonts, the StackMix method of gluing words from symbols, and the use of a generative adversarial network are explored.
Abstract
Automatic handwriting recognition is an important component in the process of electronic documents analysis, but its solution is still far from ideal. One of the main reasons for the complexity of Russian handwriting recognition is the insufficient amount of data used to train recognition models. Moreover, for the Russian language the problem is more acute and is exacerbated by a large variety of complex handwriting. This paper explores the impact of various methods of generating additional training datasets on the quality of recognition models: the method based on handwritten fonts, the StackMix method of gluing words from symbols, and the use of a generative adversarial network. A font-based method for creating images of handwritten text in Russian has been developed and described in this work. In addition, an algorithm for the formation of a new Cyrillic handwritten font based on the existing images of handwritten characters is proposed. The effectiveness of the developed method was tested using experiments that were carried out on two publicly available Cyrillic datasets using two different recognition models. The results of the experiments showed that the developed method for generating images made it possible to increase the accuracy of handwriting recognition by an average of 6%, which is comparable to the results of other more complex methods. The source code of the experiments, the proposed method, as well as the datasets generated during the experiments are posted in the public domain and are ready for download.