There are many methods to do this. as you said, you had do a transform to make sure the image in standard direction. 

if your image are in the standard scale, you can use a template match. but if has different scales, it will not work. 

but I think you can do a threshold, then binary fill, then the [no, si] will be a solid rectangle. you can do a region analysis http://scikit-image.org/docs/stable/api/skimage.measure.html#regionprops to make sure if it is your goal (by the area, perimeter and moments).  I think it will be ok!

you can also try the adboost,  sift, and other local descriptor, and so on. but I think it is too fussy and waste to solve this problem.


----- 原始邮件 -----
发件人:Fabio <fdalforno@gmail.com>
主题:[scikit-image] Fwd: Extract a particular area in document
日期:2017年05月26日 14点42分

Hi all, my name's Fabio Dal Forno.
I'm a basic scikit user and i have many question too.

I've a lot of scanned document like this

And i want to extract a particular area (si no, yes or no in english) see image below

After that i want to train a neural networ for detect the data compiled by people, y,x coordinate for detect the means of the area (materials damage ecc ecc)
and a fixed dimension image for a neural network

I started to work with the following step:

1) Homography for detect document orientation after that i transform the document for rectify that
2) Homography for detect si/no area and coordinate once

the first step work wery well but the second doesn't detect anything.

Am i wrong reasoning ?

Thanks a lot and sorry for my beginner english :)

scikit-image mailing list