2007年4月4日星期三

Testing Tesseract-1.0.3

有关tesseract的报道不再多述,google一下就会有很多结果。早在去年12月就因为要做数字识别而尝试过tesseract-1.0.2,2月底google在sourceforge上又发布了1.0.3,修正了一些内存泄露bug,并作了不少改进,至少编译起来很少了错误(vc2005对一些语法检查比较严,有一些类型转换的错误)。
1.0.3中运动WinMain,但还是通过间接调用main,完成识别工作,也就是在控制台下工作。觉得不太方便,今天花了点时间在已有的基础上增加了剪贴板功能,直接对文字识别,其中做了些彩色变换,因为目前tesseract只支持灰度图,只能识别英文和数字。下面是截图:

Our generalized Rough algorithm uses edge inlors
matron lu dehne a mapping [rot-n the orientation nfan edge paint to a reference point of the shape. Tht: reference point may be though! of as the origin of a loualeo-ordinate system for the shape Then there is an easy way ofcompuling a measure which rates how well points in the image are likely to be ortgins of the specified shape. Figure X shows a few graphic examples of the information used by the generalized Rough
transform. Lines indicate gradient directions A feature of the transform is that it will work even when the boundary is disconnected due to noise or occlusions.This is generally not true [or other strategia whichtrack edge segments
The original algonthrn by Houghm did not use

准确率还是挺高的。
目前还没有关于tesseract的官方文章,相信google不久会整理并发布的。现在也没有时间研究它,先好好把Duda的pattern classification看完再说.

further more, perhaps I've found an interesting application of tesseract, that is Visual Reader on our PC. more details, read web pages, papers, books, etc for us. This is obvious possible with tesseract and MS's Speech SDK, plus some additional image processing, for example, selecting the text region, finding out the time to perform OCR on current screen content. I've got a picture about this demo, just some programming work remained:)
life should be ease for us slothful man ~~

没有评论: