logo

• Home
• Gallery
• Questions
• Technology
• Products
• Contact

Copyright © 2008 Quixate Limited.

Questions

I already have an OCR (optical character recognition) program. Why do I need another one?

Our system is rather different from most OCR programs you might have seen.

A typical OCR system is designed to work from clearly-printed pages of text that have been carefully scanned in to your computer at high resolution. Our system is designed to work with photographs of natural scenes taken by an ordinary digital camera.

typical scanned text
Typical scanned text
 
typical photographed text
Typical photographed text

What’s so difficult about reading text from a photograph?

The text in a photograph is much harder for a computer to read than a clear scanned page of text.

The examples here show how photographed text typically suffers from poor resolution and contrast as well as compression artefacts. These effects make it much harder to separate the characters from one another to read the text.

In a scanned page, furthermore, the lines of text are straight and horizontal or almost horizontal, the fonts are consistent and of reasonable size, and the images themselves do not suffer from the distortion introduced by a camera lens, by perspective, or by uneven lighting.

In a typical image from a digital camera, however, text occurs in a wide range of fonts and at all orientations. The baseline of a row of text is typically far from straight, and the images are full of the distortions introduced by the lens of the camera and by perspective. Lighting is frequently poor. Our system copes with all these variables.

Photographic images are often full of distractions. Although a scanned page may include an image or other non-text items, these often fit tidily with the flow of the text and are relatively easy to find and ignore. Photographs, however, can be full of details — brickwork and foliage cause particular problems — that, to a computer at least, look superficially like text. Sorting out these details is well beyond the capabilities of most OCR systems.

How is your system better than the competition?

There are a few other systems available that claim to be able to read text from photographs. However, all the systems of which we are aware suffer from at least two of the following weaknesses:

  • requiring considerable manual intervention to achieve satisfactory results;
  • requiring the text to be large (many tens or even hundreds of pixels high);
  • ignoring digits, punctuation and symbols, and failing to distinguish capitals from lower-case letters;
  • having a fixed and relatively limited vocabulary, failing completely when presented with an out-of-vocabulary word;
  • being unable to correct for perspective in an image;
  • being unable to separate characters that have merged into one another as a result of blur or poor camera resolution.

Our system is fully automatic, copes comfortably with text down to eight pixels high or in some cases even smaller, and can read arbitrary text including digits and punctuation.

Can you read French/German/Spanish/... ?

At the moment our system only reads English, but the only part of it that is specific to the English language is what we call its ‘word model’. This can easily be rebuilt to suit other languages or combinations thereof. If a language uses a non-Roman script, such as Greek or Russian, the ‘character model’ also needs to be rebuilt. Please contact us to discuss your requirements if you are interested in these possibilities.

Can you read Chinese/Japanese/... ?

At the moment our system is not suitable for use with languages that have a very large number of different characters.

Can you read CAPTCHAs?

Not if your aim thereby is to send spam.