FineReader Inside
September 16, 2002

Editor's Note: FineReader is the Premium OCR engine used in the JRAPublish product.

FineReader inside
Brief description of IPA-technology and Multilevel Document Analysis

FineReader is the first OCR system of an entirely new generation substantially different from all its predecessors. The object of recognition for FineReader is the entire document and not just printed text. This has become possible thanks to ABBYYs sophisticated recognition technology based on the principles of Integrity, Purposefulness and Adaptivity (IPA Technology).

The principle of Integrity means that FineReader treats the document as a single whole consisting of integrated parts, including columns, tables, pictures, and other elements. Each of these parts is further analysed based on the same principle of Integrity, i.e. FineReader treats them as entities consisting of even smaller integrated parts. For example, a table will be perceived as an arrangement of columns, rows, and aggregated cells, while a text column will be perceived as an arrangement of lines. Thus, the analysis will continue until the smallest text elements characters are singled out.

The principle of Purposefulness means that FineReader purposefully generates hypotheses attempting to single out separate yet integrated elements on the document. To generate hypotheses, the program falls back on its years of recognition experience, or, in other words, the program behaves like a human being: it remembers the descriptions of various objects and uses these descriptions to find a matching object. It should be noted that the programs memory stores tens of thousands of such models, whereas the precise descriptions of the models allow FineReader to select the most probable models when it generates hypotheses.

Selecting the most valid and probable hypotheses is the hardest task of all. To weigh and verify its predictions, FineReader uses a special kind of Multilevel Document Analysis involving information about the objects elements, its neighbours, and superordinates. For example, to recognize a word, the program must analyse a range of possible guesses about the potential characters in this word, use morphological analysis to evaluate the likelihood of the word being correct, and not to lose sight of the broader picture, i.e. the way the whole line is broken down into words. Similarly, to recognise a table, even if there are no apparent separators in the given fragment, the program must analyse the positional relationship of the potential cells and determine the outline of the table by analysing the structure of the surrounding objects.

The principle of Adaptivity means that FineReader can adapt itself to a given document: when performing Multilevel Document Analysis, the program can use previously acquired information about the recognised text (fonts, printing faults, recurring artefacts, etc) to facilitate recognition. In other words, the program is adaptive and capable of self-learning.

Therefore, it can truly be said that FineReader is an ingenious imitation of the human intellect as applied to optical recognition since it relies on the same principles of integrity, purposefulness and adaptivity as used by most living beings. Humans and animals are champions of recognition and Fine Reader is the leading OCR system whose leadership has been recognized in numerous comparative tests and surveys.

Hosted by uCoz