Problems of Combining Multiple Text Recognition Results

Abstract

In this paper, the task of combining recognition results from multiple images is considered. Systems in which such problems occur are analyzed, and some known approaches are described. It should be noted that currently there is no unified approach that could be used to solve the combination problem for increasing text recognition accuracy using multiple images or in a video stream. As an example, a comparative study of three different approaches to the combination of per-frame recognition results of identity document fields is presented, and it is demonstrated that different approaches may be advantageous for different parts of a data set, while a selection of the potential best single result still significantly outperforms all of the analyzed methods. The task of the per-frame combination of recognition results is an important component in video stream recognition systems and requires careful consideration and the formulation of general approaches that would be applicable to various domains.