After the document analysis step, which provided informaton about the document layout, detected areas of text on each page, divided text areas into lines and these lines into wods, the individual characters (letters, number, punctuation, special signs) are detected (singled out) and can be processed.
To provide recognition results of highest accuracy, ABBYY technologies leverage advanced algorithms based on artificial intelligence as well as a broad set of classifiers (= special mechanisms for character recognition).
Following types of classifiers are used in ABBYY OCR technology:
|This classifier compares the character image with a set of pattern images. A pattern is a specially prepared image that combines all the possible ways of writing a given character. Patterns are created by superimposing many images of the same letter written in different ways. Depending on which pattern image is the best match for a given character image, ABBYY OCR technology will advance its hypotheses. The raster classifier works fast, but it cannot provide the best level of accuracy. However, it is widely used in many modern OCR programs.|
|Like the raster classifier, the feature classifier advances its hypotheses by comparing character images with pattern images. It calculates certain features of the characters under scrutiny, such as their perimeter, number of black dots in certain areas or along certain lines, etc. It is widely used in OCR programs. In certain conditions it will work just as fast as the raster classifier. Its accuracy largely depends on the right choice of features for each character (the selected features must provide necessary and sufficient information about the shape of a letter). This selection is defined by ABBYY specialists and cannot be influenced in the final product.|
|This is a sub-category of the feature classifier. It differs from the feature classifier as it calculates the features of the character's contour (and not of the entire character). This fast-working classifier is used for decorative fonts (e.g. Gothic script, Old Russian, etc.).|
|This is one of the break-through technologies developed by ABBYY. It was developed and used for recognizing hand-written characters, but it is used for mahine-printed texts as well. This classifier analyses the structure of a character by decomposing it into constituent components (lines, arcs, circles, and dots) and recreating the exact structure of the analysed character. This structural description is then compared with pattern structures. The structure classifier works slower than the raster classifier or the feature classifier but provides excellent recognition accuracy. This classifier can “restore” the missing or blotted out parts of characters.|
Feature differentiating classifier
|This classifier is used for distinguishing between very similar objects, for example the letter “m” and the ligature “rn”. This classifier is fundamentally different from the classifiers described above as it does not analyse the entire image.
Instead, it investigates only those parts of the image which may hold a clue to the solution (remember the “tree or river” example above). In the case of “m” and “n”, the clue is to be found in that part of the image where the candidate characters touch upon each other: if there is a gap between them and this gap is wide enough, the program 'understands' that it is dealing with two separate characters.
Image Description: The feature differentiating classifier at work. To choose between two similar characters (e.g. between the letters “D” and “O”), the classifier calculates feature parameters (the slope of a line in a crucial area).
Structure differentiating classifier
|This is a very accurate classifier which might even be called the 'heavy artillery of OCR'. Like the structure classifier, it was initially developed and used for recognizing hand-written texts. It is also used for distinguishing between very similar characters but is a lot more precise as it analyses the structure of characters. It can recognize even distorted letters.|
Image Description: The feature differentiating classifier at work. To choose between similar characters (e.g. between the combination “f1” and the letter “A”, the classifier compares the structure of the letters paying particular attention to their contours.