Poker-AI.org
http://poker-ai.org/phpbb/

OCR for asian network
http://poker-ai.org/phpbb/viewtopic.php?f=26&t=3027		 Page 1 of 1
Author:  mlatinjo [ Sat Mar 25, 2017 5:53 pm ]
Post subject:  OCR for asian network
Hi,

I am trying to do ocr for asian network, it seems very complicated so far because letters, digits, cards are displayed in many different ways. I can separate digits / letters automatically and save them as templates, but the problem is that I don't know if there are limited number of templates for e.g. digit 0.
in the attachment you can find 3 different stacks, so you can see that the same digits look different. There are occasions when the same digit (e.g. 0) looks the same way, but most of the time it is different. I convert the image to binary in order to remove rgb but still there are a lot of differences. You can also see 7 different digits 0
that are convert to binary image, how it all looks very different.

I haven't tried tesseract on this, but i think that it would perform very bad because letters and digits are very often connected, no clear border.
I hope that you will recognize that there are two images in the attachment, the second one displaying 3 different stacks is very small.
Any idea how to solve this?
Thanks in advance.

Attachments:
tables.png
tables.png [ 975.93 KiB | Viewed 25243 times ]
image.png
image.png [ 22.9 KiB | Viewed 25246 times ]
stacks.png
stacks.png [ 1.55 KiB | Viewed 25246 times ]
Author:  HontoNiBaka [ Sat Mar 25, 2017 10:42 pm ]
Post subject:  Re: OCR for asian network
What is the name of the site? Maybe you can change the fonts or so, to make it more readable?

Everytime I did OCR, I was never happy with software like Tess, I always wrote my own code. I also never use anything like NNs in my code, I usually write a ton of custom rules to detect the borders and the characters. It's usually a ton of work, but after a while the accuracy is pretty good.
Author:  nefton [ Sun Mar 26, 2017 1:01 am ]
Post subject:  Re: OCR for asian network
Today I founded some new filters for this situation.
c++ code (opencv)
Code:
cv::Mat Img = cv::imread("c:/orign.png");
   
   //filter1
   for (int row = 0; row < Img.rows; row++){
      for (int col = 0; col < Img.cols-1; col++){
         uint8_t b1 = Img.at<cv::Vec3b>(row, col)[0];
         uint8_t g1 = Img.at<cv::Vec3b>(row, col)[1];
         uint8_t r1 = Img.at<cv::Vec3b>(row, col)[2];
         uint8_t h1, s1, v1;
         RGB2HSV(r1, g1, b1, &h1, &s1, &v1);
         uint8_t b2 = Img.at<cv::Vec3b>(row, col+1)[0];
         uint8_t g2 = Img.at<cv::Vec3b>(row, col+1)[1];
         uint8_t r2 = Img.at<cv::Vec3b>(row, col+1)[2];
         uint8_t h2, s2, v2;
         RGB2HSV(r2, g2, b2, &h2, &s2, &v2);
         if ((h1 > h2)&&(h1-h2 > 15)&&(h1>100)&&(h1<125)) Img.at<cv::Vec3b>(row, col+1) = cv::Vec3b(0, 0, 0);
      }
   }
   cv::imwrite("c:/filter1.png", Img);
   //filter2
   for (int row = 0; row < Img.rows; row++){
      for (int col = 0; col < Img.cols - 1; col++){
         uint8_t b1 = Img.at<cv::Vec3b>(row, col)[0];
         uint8_t g1 = Img.at<cv::Vec3b>(row, col)[1];
         uint8_t r1 = Img.at<cv::Vec3b>(row, col)[2];
         uint8_t h1, s1, v1;
         RGB2HSV(r1, g1, b1, &h1, &s1, &v1);
         uint8_t b2 = Img.at<cv::Vec3b>(row, col + 1)[0];
         uint8_t g2 = Img.at<cv::Vec3b>(row, col + 1)[1];
         uint8_t r2 = Img.at<cv::Vec3b>(row, col + 1)[2];
         uint8_t h2, s2, v2;
         RGB2HSV(r2, g2, b2, &h2, &s2, &v2);
         if ((h1 > h2) && (h1 - h2 > 15) && (h2>100) && (h2<125)) Img.at<cv::Vec3b>(row, col) = cv::Vec3b(0, 0, 0);
      }
   }
   cv::imwrite("c:/filter2.png", Img);
   //cv::Mat Img2 = Img.clone();
   //filter3
   for (int row = 0; row < Img.rows; row++){
      for (int col = 0; col < Img.cols; col++){
         uint8_t b = Img.at<cv::Vec3b>(row, col)[0];
         uint8_t g = Img.at<cv::Vec3b>(row, col)[1];
         uint8_t r = Img.at<cv::Vec3b>(row, col)[2];
         uint8_t h, s, v;
         RGB2HSV(r, g, b, &h, &s, &v);
         bool is_good_point = false;
         if ((h>100) && (h < 125) && (s>60) && (s<120) && (v > 200)) is_good_point = true;
         
         if (!is_good_point) Img.at<cv::Vec3b>(row, col) = cv::Vec3b(0, 0, 0);
         
      }
   }
   cv::imwrite("c:/filter3.png", Img);


Attachments:
filter2.png
filter2.png [ 25.53 KiB | Viewed 25233 times ]
filter1.png
filter1.png [ 26.28 KiB | Viewed 25233 times ]
orign.png
orign.png [ 14.96 KiB | Viewed 25233 times ]
Author:  nefton [ Sun Mar 26, 2017 1:02 am ]
Post subject:  Re: OCR for asian network
Cant attach more 3 files. And here is result.
(you can do all 3 filters in one sycle)

With good filters, it will be much easier to recognize ;)

Attachments:
result.png
result.png [ 12.33 KiB | Viewed 25233 times ]
Author:  nefton [ Sun Mar 26, 2017 1:19 am ]
Post subject:  Re: OCR for asian network
And with close letters.
If letter you cut from stack is too width (9+ pixels)
You can try to compare all leters you have with the begin of your superwidth twice letter,
and if sucsess try the same thing on begin+sucsess_letter.width position
Author:  nefton [ Sun Mar 26, 2017 12:04 pm ]
Post subject:  Re: OCR for asian network
Can you give me the biggest size stack you can do?
Have some idea.
Author:  mlatinjo [ Sun Mar 26, 2017 4:21 pm ]
Post subject:  Re: OCR for asian network
Hi thank you for your answer. I don't understand what kind of filters do you apply on the image? Could you describe it in words? I am using c#, for my bot, I understand c++ code but I don't understand what you are doing exactly? I see that you change image from RGB to HSV but that is all I understand.
Also by looking at your output images I see that the digits are still quite different, so how can I recognize e.g. digit 0 if it is displayed in so many different ways?
In the attachment you can find the biggest stack size I have found.

Attachments:
table_0.png
table_0.png [ 401.06 KiB | Viewed 25216 times ]
Author:  mlatinjo [ Sun Mar 26, 2017 4:23 pm ]
Post subject:  Re: OCR for asian network
nefton wrote:
Cant attach more 3 files. And here is result.
(you can do all 3 filters in one sycle)

With good filters, it will be much easier to recognize ;)


Ok I can see from the result image that it should be easier to recognize digits after the proposed filters are applied. Still I don't understand how could I do it in c# because I don't understand your code?
Author:  mlatinjo [ Sun Mar 26, 2017 4:24 pm ]
Post subject:  Re: OCR for asian network
HontoNiBaka wrote:
What is the name of the site? Maybe you can change the fonts or so, to make it more readable?

Everytime I did OCR, I was never happy with software like Tess, I always wrote my own code. I also never use anything like NNs in my code, I usually write a ton of custom rules to detect the borders and the characters. It's usually a ton of work, but after a while the accuracy is pretty good.


The name is natural 8, asian online casino.
Author:  nefton [ Sun Mar 26, 2017 5:32 pm ]
Post subject:  Re: OCR for asian network
mlatinjo wrote:
In the attachment you can find the biggest stack size I have found.

I mean width and height of stack digits. (Screenshoot max table size you can do)

Filters... I thought the above pictures will be enough for understanding.
1st filter. I delete right edge of all digits. I see that on edge there is big hue gradient (h1-h2)
2nd filter. The same only on right edge
3rd. remain only desired hue and saturation diapason.

Without 1st and 2nd filters - there was bad results from 3rd filter.

But I have another idea how to recognize it. May be this red and blue edges are not noise, and we should not remove it.
They appear from worst algoritm of resizing they template digits to size that are displayed to you.
And if we can understand this algoritm of resizing - we can use this bad pixels to restore the original digits.

And even if this idea will not work - you can just store many variants of your digits (10variants x 10digits = 100 templates only (it is not terrible value))
Author:  mlatinjo [ Sun Mar 26, 2017 6:19 pm ]
Post subject:  Re: OCR for asian network
Hi in the attachment you can find the biggest size table I could make for my 22 inch monitor, and standard table size. I guess that the resize is made on standard table size, which is what you want to see?
You can see that even in the standard table size, the digits look quite different. From that I can conclude that the digits and letters don't look different because of resize because on standard table size it also looks all different?

Attachments:
standardSizeTable.png
standardSizeTable.png [ 512.74 KiB | Viewed 25210 times ]
bigtable.png
bigtable.png [ 1.19 MiB | Viewed 25210 times ]
Author:  nefton [ Sun Mar 26, 2017 8:27 pm ]
Post subject:  Re: OCR for asian network
I do it :) !!!!
1. This color artifacts calls clear type!
2. Here is macro photo of orign.png on my monitor. It all explain.
Attachment:
clear_type.jpg
clear_type.jpg [ 592.89 KiB | Viewed 25206 times ]

red lines - is border of monitor pixels.
So now I think it easy to raise qualyty of recognizing X10 times :idea: :geek:
Author:  mlatinjo [ Mon Mar 27, 2017 2:41 am ]
Post subject:  Re: OCR for asian network
How did you produce image1? It is interesting how you can see rgb channels of each pixel in the monitor.

How does scaling to huge image size is going to help recognize digits?
Author:  nefton [ Mon Mar 27, 2017 3:19 am ]
Post subject:  Re: OCR for asian network
[quote="mlatinjo"]How did you produce image1? It is interesting how you can see rgb channels of each pixel in the monitor.
Attachment:
macro.jpg
macro.jpg [ 92.32 KiB | Viewed 25198 times ]

small hole on the cap is to make minimum apperture I can do. (for big depht of field)
+ several macro rings M42
With this tehniques I can do 5x times more closer picture then in photo below. And more quality :roll:
This photo makes in hurry.
Author:  nefton [ Mon Mar 27, 2017 3:27 am ]
Post subject:  Re: OCR for asian network
mlatinjo wrote:
How does scaling to huge image size is going to help recognize digits?


No, understanding of what happens will help mutch.
You can imagine several algoritms, based on subpixel places, and all of them will be mutch better then now.
Simplest - is to make width*3, height*3 1 chanel picture and fill it by values of r, g, b subpixels.
I tryed it - it good.
But it is not best algoritm for you. More understanding cleartype algoritm will produce mutch better algoritms.(but some more hard)
Author:  mlatinjo [ Mon Mar 27, 2017 6:42 pm ]
Post subject:  Re: OCR for asian network
Quote:
Simplest - is to make width*3, height*3 1 chanel picture and fill it by values of r, g, b subpixels.
I tryed it - it good.


Just to mention, it is also important to recognize table information fast, it would be good less that 0.5 sec. I think that if the solution would be to increase image then it would increase the time to recognize information.
Author:  nefton [ Mon Mar 27, 2017 8:44 pm ]
Post subject:  Re: OCR for asian network
No, we talk about cutted stack image (about 50x10) and no whole image ofcource.
It is nothing. You can not even measure this time without many itterations.
Anyway, all algoritms I thinking includes 1st step triple at least width.
And for debug sences you should triple height too. (after you had compleet working algoritm you can reduce height trippling)
Page 1 of 1 All times are UTC
Powered by phpBB® Forum Software © phpBB Group
http://www.phpbb.com/