Flexible optimization of text recognition algorithms

This paper presents a system for the optimization of text recognition algorithms. First a theoretic four-staged model of text recognition is proposed. In this four-staged model, the second stage called text localization is optimized. A reinterpreted version of the F measure is used as a fitness indicator for optimization of the localization. The optimization method is described and the role of the algorithm of Nelder and Mead in the optimization process is explained. The system is introduced and it is indicated, how it can be extended with custom algorithms. Selected experimental results are presented at the end of this work. The optimization approach can improve existing text localization algorithms on untrained data up to 87% of their base localization rate in F measure category.