I think Dmitri's suggest to start simple is a good one, but, if you need
it, don't forget that you've got a lot of other information that can be
leveraged to help. The notes all have a fixed aspect ratio (and size?).
They've got a relatively standard layout. The denomination is encoded
multiple places on the note. You can use as much of this additional
information as you need to help make the solution easier (or to cross check
the quality of your result).
Tom
On Friday, October 2, 2015 at 4:39:12 PM UTC-4, Dmitri Silaev wrote:
>
> Hi Juan Pablo,
>
> Here are my thoughts about how I'd go with the initial version of image
> processor. I'm not sure OpenCV is the best tool for doing all this and
> you're free to choose how to implement this.
>
> As far as I understand the idea of your app, bills can be captured in an
> arbitrary manner, but in most cases close to horizontal position and at a
> projection close to fronto-parallel. This allows us to assume that images
> would have minimal distortions of characters and symbols. Also, the user
> would try to capture as much bill area as possible, hence another
> assumption - most of the white area containing tactile symbols would be
> visible in the captured image.
>
> That said, for the initial version I'd not even bother with locating the
> full bill area, i.e. edge detection, etc. but only would do a simple search
> for the biggest and lightest area. That should be the "white area". In this
> area, apply a simple threshold and you'll get connected components (CCs).
> Among them, you'd need only those that are close in form to rotated
> squares. They are also easy to find as such CCs fill up almost entire area
> of their rotated bounding rectangle. These CCs would be those tactile
> symbols you are after.
>
> The rest is trivial - count tactile symbols and get the denomination of
> your bill.
>
> Of course, you'd add more sophistication to cope with real world images
> but the backbone of the algorithm looks to me like this. All work is done
> in grayscale.
>
> HTH
>
> Best regards,
> Dmitri Silaev
> www.CustomOCR.com
>
>
>
>
>
> On Mon, Sep 28, 2015 at 12:18 AM, Juan Pablo Aveggio <jpav...@gmail.com
> <javascript:>> wrote:
>
> Hi Dmitri Silaev
>
> Thanks for your useful help. Actually I have almost no progress, in terms
> of image preprocessing. Just convert the image to grayscale before applying
> OCR. But I could not get good training data. The test code is as follows:
> #include <opencv2/highgui/highgui.hpp>
> #include <opencv2/imgproc/imgproc.hpp>
> #include <tesseract/baseapi.h>
> #include <leptonica/allheaders.h>
> #include <iostream>
>
> using namespace cv;
>
> int main(int, char**)
> {
> VideoCapture cap(0); // open the default camera
> if(!cap.isOpened()) // check if we succeeded
> return -1;
> int c;
> Mat gray;
> namedWindow("gray", 1);
> tesseract::TessBaseAPI tess;
> tess.Init("/usr/share/tesseract-ocr/tessdata/", "bil",
> tesseract::OEM_DEFAULT );
> tess.SetPageSegMode(tesseract::PSM_SINGLE_WORD);
>
> for(;;)
> {
> Mat frame;
> cap >> frame;
> cvtColor(frame, gray, CV_BGR2GRAY);
>
> c = waitKey(30);
> if(c == 27) break;
> else if(c > 0) {
> tess.SetImage((uchar*)gray.data, gray.cols, gray.rows, 1,
> gray.cols);
> Boxa* boxes = tess.GetComponentImages(tesseract::RIL_WORD,
> true, NULL, NULL);
> for(int i=0; i < boxes->n; i++){
> BOX* box = boxaGetBox(boxes, i, L_CLONE);
> rectangle(gray, Point(box->x, box->y),
> Point(box->x+box->w, box->y+box->h), Scalar(255, 0, 0));
> }
> char* out = tess.GetUTF8Text();
> std::cout << out << std::endl;
> imshow("gray", gray);
> waitKey(4000);
> }else imshow("gray", gray);
> }
> tess.~TessBaseAPI();
> // the camera will be deinitialized automatically in VideoCapture
> destructor
> return 0;
> }
>
> With this code and my training data only thing I've done is draw a square
> around the pattern, but only when I place the ticket close enough to the
> camera and exactly horizontal. It has yielded some results, but with very
> low hit. I also tested PSM_SINGLE_CHARACTER page segmentation mode, with
> similar results.
> I thought all this could be due to errors thrown during training process,
> which resulted in bad training data. Now I understand that this is because
> my characters appear disconnected, isolated, and tesseract is designed to
> detect horizontal lines of text with words, mostly several characters.
> Then I could just use OpenCV to solve this problem? The hardest part seems
> to be finding the region where the pattern in the bill, and its rotation
> is. Once echos with this information, I could straighten out and deal with
> this small subimage more easily.
> I do not have much experience with OpenCV. But I'm willing to learn. I
> imagine that we will have to apply an algorithm to detect edges or corners,
> to try to get the contour of the ticket. We have to consider that the
> ticket might be being partially captured on camera. It could even be the
> reverse, so that the pattern will not be found. I see it quite difficult,
> but it's a good challenge.
>
> Finally, I would note that I have selected this pattern because I thought
> it would be easier to detect. They are also issuing a new currency, with
> many different typefaces and design, but the pattern have not changed. But
> any suggestions are welcome.
>
> Thank you very much for your interest.
> Best regards
> Juan Pablo Aveggio
>
>
> El sábado, 26 de septiembre de 2015, 11:36:18 (UTC-3), Dmitri Silaev
> escribió:
>
> Hi Juan Pablo,
>
> The problem cannot be solved by Tesseract as is. Even given such perfect
> images like you've shown, Tesseract would fail since your "characters" are
> too disjointed, have no meaningful baseline and only happen as singletons.
>
> However a simple and robust recognition can be implemented without
> Tesseract using common sense and a bit of programming. Of image processing
> operations, you only would need trivial thresholding. Though, some more
> involved image preprocessing is required to convert the image to the form
> close to what you've demonstrated in your sample images.
>
> The said preprocessing would be needed anyway even if Tesseract worked for
> your "characters". Tell what you already have done so far in this direction
> so I can share more details about the above method, if you wish.
>
> -Dmitri
> Hi Dmitri Silaev.
>
> Thanks for reply. They are bills, sorry for mistranslation. You can see
> examples:
> 2 <http://k43.kn3.net/AE7DA8C86.jpg> 5 <http://k36.kn3.net/43DDCD402.jpg>
> 10
> <http://radioalfa971.com.ar.elserver.com/wp-content/uploads/2013/12/Billete-Argentino-10-pesos.jpg>
> 20 50
> <http://rafaelanoticias.com/privados/subidas/fotos/3783_billete_de_cincuenta_50_pesos.jpg>
> 100 <http://k02.kn3.net/58A05701F.jpg>
>
> These patterns have relief for the blind, but they are very worn and no
> longer apply. So I'm working on an android app to detect the value
> and speech it to user.
>
>
> El miércoles, 23 de septiembre de 2015, 18:26:43 (UTC-3), Dmitri Silaev
> escribió:
>
> Hi Juan Pablo,
>
> The problem seems interesting. However not sure if you can use Tesseract
> for that. Could you show one or more example tickets?
>
> Best regards,
> Dmitri Silaev
> www.CustomOCR.com
>
>
>
>
>
> On Tue, Sep 22, 2015 at 2:17 AM, Juan Pablo Aveggio <jpav...@gmail.com>
> wrote:
>
> Hello
> I'm trying to train tesseract for recognition of patterns present in
> tickets. Each ticket possesses a unique pattern in a predetermined place
> which determines its value. As these patterns are not including unicode
> characters, I assigned them the characters 'a' to 'f'.
> I created a .tif image with six patterns:
> bil.pat.exp0.tif
> <https://drive.google.com/file/d/0B7CfYFzWHQDAYWU4M3hIQXUyOWs/view?usp=sharing>
> and the corresponding file box:
> bil.pat.exp0.box
> <https://drive.google.com/file/d/0B7CfYFzWHQDAVkJlZ3lreEdpaXc/view?usp=sharing>
> a 32 692 165 958 0
> b 221 734 354 958 0
> c 32 446 165 628 0
> d 221 488 354 628 0
> e 32 275 165 373 0
> f 221 317 277 373 0
>
> Then I ran:
> tesseract bil.pat.exp0.tif bil.pat.exp0 box.train
> and output:
> Tesseract Open Source OCR Engine v3.04.00 with Leptonica
> Page 1
> APPLY_BOXES:
> Boxes read from boxfile: 6
> APPLY_BOXES: Unlabelled word at :Bounding box=(-958,221)->(-734,277)
> APPLY_BOXES: Unlabelled word at :Bounding box=(-628,221)->(-488,277)
> APPLY_BOXES: Unlabelled word at :Bounding box=(-958,32)->(-734,88)
> APPLY_BOXES: Unlabelled word at :Bounding box=(-628,32)->(-488,88)
> APPLY_BOXES: Unlabelled word at :Bounding box=(-373,32)->(-317,88)
> Found 6 good blobs.
> 5 remaining unlabelled words deleted.
> Generated training data for 6 words
> That can not mean negative coordinates. Despite this I tried to keep going.
> My font_properties is:
> bil.pat.box 0 0 1 0 0
> bil.words_list is:
> a
> b
> c
> d
> e
> f
>
> then I ran:
> $ unicharset_extractor bil.pat.exp0.box
> Extracting unicharset from bil.pat.exp0.box
> Wrote unicharset file ./unicharset.
> but the unicharset file has:
> 9
> NULL 0 NULL 0
> Joined 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # Joined [4a 6f 69 6e 65
> 64 ]
> |Broken|0|1 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # Broken
> a 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # a [61 ]
> b 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # b [62 ]
> c 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # c [63 ]
> d 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # d [64 ]
> e 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # e [65 ]
> f 0 0,255,0,255,0,0,0,0,0,0 NULL 0 0 0 # f [66 ]
> Then I ran:
> $ mftraining -F font_properties -U unicharset -O bil.unicharset bil.pat.
> exp0.tr
> Read shape table shapetable of 0 shapes
> Reading bil.pat.exp0.tr ...
> Bad properties for index 3, char a: 0,255 0,255 0,0 0,0 0,0
> Bad properties for index 4, char b: 0,255 0,255 0,0 0,0 0,0
> Bad properties for index 5, char c: 0,255 0
>
> ...
--
You received this message because you are subscribed to the Google Groups
"tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to tesseract-ocr+unsubscr...@googlegroups.com.
To post to this group, send email to tesseract-ocr@googlegroups.com.
Visit this group at http://groups.google.com/group/tesseract-ocr.
To view this discussion on the web visit
https://groups.google.com/d/msgid/tesseract-ocr/f51aef72-5fd7-4f82-80fa-36bb268abbb3%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
