Thursday, June 21, 2007
From time to time, our own T.V. Raman shares his tips on how to use Google from his perspective as a technologist who cannot see -- tips that sighted people, among others, may also find useful. - Ed.
As someone who cannot see, I prefer to live in a mostly paperless world. This means ruthlessly turning every piece of paper that enters my life into a set of bits that I can process digitally. I scan in everything. Until now, I have relied on commercial OCR packages to convert these images into readable text. OCR is perhaps one of the areas where the benefits of Moore's Law are most evident; today, OCR can do remarkably well when handed a page image. Until now, my only dissatisfaction with the status quo in this area has been that commercial OCR engines afford me little flexibility with respect to training them to do better on documents that are specific to me.
The advent of our own open source OCR initiative, OCRopus (source code: Ocropus Sources) is a welcome change in this regard. I introduced support for OCRopus in Emacspeak recently, and the HTML output this produces compares favorably with output from commercial OCR engines, provided you place the page at the right orientation on the scanner. OCRopus' extensibility, and the ability to express the OCR as a structured HTML document makes it an ideal starting point for producing rich spoken output. The possibilities are enormous for people being able to collectively train, customize and improve an OCR engine.