Meeting interesting characters

(Document analysis at DFKI - Thomas Pötter)

Summary

The presentation deals with different methods for optical characters recognition (OCR). The most important methods are:

• Shape-oriented recognition: The inner and outer boundaries of characters are detected and saved as vectors or splines. This detection method requires several passes with relaxations in order to eliminate accidental roughness due to errors in the scanning process or due to letters of bad quality in the original document. On the other hand it mustn't be relaxed so far that important features of character disappear and are lost for the further recognition process. This method is particularly useful for recognizing characters at different sizes and for new characters because it captures all the character's characteristics -- on the other hand it is slow and hard to implement.
• Crossing counts ("Winkelschnittanalyse"): The character is intersected with several parallel lines. For each line the number of intersections is recorded. The intersections are calculated with parallel lines from/at(?) different angles (usually 8 different angles). This method is faster and can recognize characters at any size as well. Italics and obliques are problematic.
• Neural network based on pixmap (perceptron): The pixel matrix of the original characters are learned by the neural network. Characters to be recognized are given to this neural network as pixmap. Noisy or blurred characters are easily recognized if most of a character's pixels remain in place. Different sizes and characters enlarged in just one direction (along x- or y-axis) have to calculated back to the original size of the pixmap. This method is easy to implement and works fine for printed characters at a uniform size only.
• Hybrid methods: Some criterion like the position of stems, round corners, line ends or data determined by the crossing counts method is taken as input for a neural network.

Hand-written text: Deskewing is required to cope with slanted handwriting and it is difficult to break up words into characters. Varying characters are the hardest problem: Notepad computers record the whole process of writing and often require that the user writes each character into a separate bounding box provided by the system to facilitate OCR.

Stages in text processing (at DFKI):

1. Segmentation: pre-classifying objects as graphic or text (lines, words, characters).
2. Recognition/Classification: Characters are recognized: For each character different possibilities are provided together with certainty factors which can be viewed as probabilities for each alternative.
3. Post-classification: Taking advantage of the graphical and textual environment to cut down the number of alternatives and to calculate better certainty factors. The graphical information used is the relative position to the halfline, capline, baseline of a character: That way ,'cCoOxXsSvVwWpP can be classified properly. Data from the textual environment is used to determine the syllables to which a character might belong, to determine the correct punctuation marks and the spacing in between the words as well as to make a dictionary lookup to make sure that all the characters of a word a recognized properly. Interesting concepts in this concept are: confusion matrices and the calculation of a smallest edit distance between the recognized word (possibly misspelled) and the words in the dictionary.
4. Text analysis phase: E. g. classification of a business letter as invoice, order, inquiry, etc. Data like sender; subject, date, ordered products, related letters are determined. This information will be stored in a database. Additional functionality is to "understand" the text and to answer queries automatically and to take further steps necessary to process orders and invoices.

Applications:

1. Archiving business letters, faxes, e-mails, etc. consistently
2. Office automation: answering inquiries automatically and doing automatic processing depending on the nature of the letter.
3. Processing filled in blanks automatically (for use in the tax office or other public bodies)
4. Classifying receipts to automate the creation of balances in business.
5. As substitute for typing written information into the computer.