Authors: Tatiana Novikova, Olga Barinova, Sergey Milyaev, Vladimir Kirichenko, Alexander Sapatov, Pushmeet Kohli, Victor Lempitsky - Lomonosov Moscow State University, Moscow, Russia

Description: Our reading system formulates the problem of word recognition as the maximum a posteriori (MAP) inference in a unified probabilistic framework [1]. Our model enforces both the language consistency, and the consistency of the attributes of letters that constitute a word. This unified treatment allows us to carry over the uncertainty associated with character detection and recognition in a principled fashion while enforcing structure and language constraints. For the MAP inference we use weighted finite-state transducers (WFSTs).We utilize different operations on WFSTs as building blocks within our inference procedure to ensure low computational cost even in the presence of very large (language-scale) lexicon priors. [1] Novikova Tatiana, Olga Barinova, Pushmeet Kohli, and Victor Lempitsky. "Large-lexicon attribute-consistent text recognition in natural images." In Computer Vision–ECCV 2012, pp. 752-765. Springer Berlin Heidelberg, 2012