Reading text is a simple enough task for humans. But unless it’s cleaned up and served on a plate computers just can’t do it.
At least they couldn’t until Mireille Boutin and pals from Purdue University took a shot at the problem.
These guys have built an impressive algorithm that looks for and finds text in real-life cluttered images.
And it works well. In their, albeit limited, tests on 65 real-life images, the algorithm correctly identified the text 97 per cent of the time.
Cars that can read signposts, anyone?
Ref: arxiv.org/abs/0801.4807: Automatic Text Area Segmentation in Natural Images