Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Офтальмология / Английские материалы / Assistive Technology for Visually Impaired and Blinde People_Hersh,Jonson_2008.pdf
Скачиваний:
0
Добавлен:
28.03.2026
Размер:
12.16 Mб
Скачать

15.2 Basics of Optical Character Recognition Technology

561

Figure 15.3. Individual character images of varying quality

that do image normalization the only difference between a period and a vertical bar (which could be the letter ‘l’, the numeral ‘1’ or the bar character itself ‘|’) is the aspect ratio calculated before fitting the character onto a normalized grid.

Sometimes the differences between different characters can be quite small. The difference in a small point-size font between a small letter ‘e’ and a ‘c’ turns on a few pixels: if the contrast is set too lightly and the font is small, these differences might not be captured. As noted above, in some fonts there are no differences in appearance between the letter ‘l’ and the digit ‘1’, or the letter ‘o’ and the digit ‘0’, so these tend to be grouped as one recognized character.

Contextual analysis is used to separate these similar looking characters. For example, “look” can be easily told to be all letters, whereas “1004” is clearly all numbers. “Thc” is probably wrong and can be changed to “The” with confidence that this is almost always correct. If the word “return” is printed with the final ‘r’ and the ‘n’ touching, it may look like the word “retum”. OCR systems use dictionary checks to see what words are permitted in a given language, turning “retum” into “return,” but this is not always possible. For example, “modem” and “modern” are both allowable words in English. So, it is usually easy to tell when a book has been scanned using OCR, because the OCR makes mistakes that humans rarely would.

Since humans can read text with many errors and still understand what the correct text should be (an example is the deliberate misspelling in many current e- mail spam messages), we know that it is possible to solve this problem better. This is an area where continuing to enhance the contextual and document understanding technology would make character recognition much more accurate.

15.2.2 Practical Issues with Optical Character Recognition Technology

Users rate OCR technology on accuracy and speed. How good a job did the OCR do in recreating the page? Ideally, this is as similar as possible to the author’s original electronic document. How long a wait is there from scanning to reading a given page? As short a wait as possible is the desirable outcome, since sighted people are used to reading instantaneously.

The most common way of measuring OCR accuracy calculates a character accuracy rate measured by the percentage of the characters that are correctly rec-