Instead of using rectangular frames for recognition, the new algorithm employs ‘tube’ representations of text reference frames.
Developers have been working on Optical Character Recognition systems (OCR) since the early 1970s. Today, this field is faced with a new set of challenges - recognizing complex fonts and handwritten text.
A group of Amazon programmers have developed a technology called TextTubes, which creates models of text in the form of tubes around the medial axis of the text.
In an article published on arXiv.org the developers explained their solution in detail. Unlike traditional techniques, the new algorithm uses ‘tube’ representations of text reference frames instead of rectangular frames.
The developers tested their technology using two datasets: CTW-1500 (1500 images) and Total-Text (approximately 1255 images). All test images contained at least one inscription written in a curved font.
The TextTubes technology outperformed traditional approaches, with 83.65% success rate versus 75.6%.
Share this with your friends!