Program > IAPR invited Talks
Three IAPR invited talks will be provided during the workshop:
Vladimir Rybkin - Mr. Anisimovich - Character Recognition and Image Processing Group - ABBYY.
An Inside Look into ABBYY OCR Technology
This keynote uncovers fundamental aspects of ABBYY FineReader OCR technology, the technology that empowers a series of ABBYY’s software products.
Ideas behind FineReader technology are holistic approach and purposeful search. In top-to-bottom approach complex object is divided into simple ones that can be easily recognized. In bottom-to-top approach easily recognizable simple objects are combined to form a complex object. Holistic approach deals both with complex object as a whole and with its more simple parts, observing relationships between these parts. It considers different variants of divisions, combinations and recognition of parts in order to recognize the whole object.
FineReader applies holistic approach on the word level, taking word image as an input and producing a list of word recognition variants as an output of this level.
Each word is treated as a complex object consisting from more simple objects called graphemes. A “grapheme” concept is introduced and used alongside with “character” concept in order to differentiate shape and semantics. Division of a word into graphemes is redundant and is targeted to produce all reasonable variants of graphemes, not just the most probable ones. Several classifiers are used to produce recognition variants for each presumable grapheme. Separate classifier is used to order these variants. Results of division and recognition are represented as a Graph of Linear Division (GLD).
The context analysis stage works on GLD and assembles word recognition variants from graphemes recognition variants. Purposeful search principle is exploited and word variants are build according to a set of rules or grammars called “models”. Models aim the search and help to avoid building word variants that are just arbitrary combinations of characters. Word variants are evaluated and ordered based on the set of features including (but not limited to) graphemes recognition quality, model type and geometrical features.
Further improvements can include promotion of holistic approach above the word level, fine-tuning using automated tools and large datasets and implementing on-the-fly adaptability.
Prof Andreas DENGEL - DFKI Kaiserslautern - Germany
I am asking this question because I am convinced that it is time to redefine the target of our studies to a more open understanding of the term document. Therefore, in my talk I like to give an analytical view to our research field and address the various dimensions of a “modern” document trying to impulse a rethinking in understanding document analysis and recognition. Further, my intension is to stimulate a discussion about what might be research focus of our community in the coming years and how to attract people who study the nature of a document from a different perspective.
This keynote will cover he following topics: