I'm facing problem of letters extraction. The image is grayscale. The letters are in a row. Background of the image is not homogenous, there can be some texture with some non-white intensity. The letters are black. And of course, the letters are sometimes so similar with the background so they cannot be separated easily. The problem is also similar to car plate letters extraction. But in these case, we can expect "damaged plate with damaged letters". Unfortunately, the particular images are subject of secret project, so I cannot include there an example.
I used several types of thresholds including adaptive ones, histograms-based methods and some segmentation-based method. But no one of them works generally good.
The goal of my task is to extract letters, i.e., detect rectangular area of each of them, or to exclude them from backgroud.
There are two criterias: success rate (I need almost 100 % success rate of extraction) and processing speed (as fast as possible, several ms ideally).
Thank you for your advices, or links to some useful papers.
There is one, by myself artificial created image. As I said, the original images cannot be published. But this example is really close to original cases. In my application I can recognize these letters, but only in the case that I can extract them. And for, e.g., letters 0 and 3 it is difficult to separate them if they are "connected".
Dear S. T. Patil,
can you formulate it in more details please? I'm not sure what should be input and learn set to some machine learning algorithm (which one?) to obtain extracted letters from background:(
Hey Petr,
as a very basic approach, given your images look like that most of the time, you could use 2D-correlation (like in older OCR systems). Build a database with example letters, correlate them and pick the ones with the highest correlation coefficient. You might have to solve some ambiguous cases (like 6 & 8 for example) afterwards.
A problem with this method is, that it is rather scale-dependent...
Greetings, David
One of the well-known answer for your problem is the use of 2D-correlation as properly suggested by David Jule Mack. In such way, algorithm can extract and detect the letters simultaneously as process goes on.
Hey David,
it's not bad idea, thank you. The problem is that I have been thinking the same way, so I tried it. It does not work, because in my case there are different scales, letters can be partially rotated, they can have different fonts... So even if I created "database" of thousands these letters, tried to match them by best corelation and take the position, there were a lot of mistakes:(
Dear
Look the link please.
https://support.office.com/en-us/article/Extract-text-from-pictures-and-file-printouts-by-using-OCR-in-OneNote-93a70a2f-ebcd-42dc-9f0b-19b09fd775b4
https://www.youtube.com/watch?v=TWJXCJA_hzI
Regards, Shafagat
Thanks everybody for help. Hopefully, the problem has solution which consists of several steps:
1, Outliers removing – i.e., remove some black homogenous area near border
2, Normalization – the image is adaptive normalized (so it definitely includes max intensity 255) and then the intensities are linearly shifted out (intensities higher than 255 are round to 255) until remains only several percent black pixels. I expect that letters take approx. 15 % of image, so I shift the intensities until only approx. 15 % remains. Then the remain pixels intensity are return to their original value (so they are black). After then, only letter remains (but they still remain connected)
3, after each letter (or group of letter) the smallest rectangular area is created
4, from the areas median of their widths is searched. Due to the median too thin areas in neighborhood are connected into one bigger and too huge areas are divided into smaller.
With this steps I’m now able to extract letters independently on the image contrast, the letters connections etc. I hope, these steps can be useful for you as well.
You can use online API http://www.ocrwebservice.com/ . they provide conversion of 25 pages per day, for free.
If this is not your core research area, then this API can save you alot of time .
good luck
Regards Murad.
Hi
I would suggest you to look at
http://www.mathworks.com/help/vision/examples/automatically-detect-and-recognize-text-in-natural-images.html
were the use MSER to get areas that are the good candidates for letters. After that you can use Connected component analysis to choose which areas are letters and which are not using the fact that all the letters are in the same row. A good way to find letters that are connected together is looking at the ratio between the width of the connected component and the height.
Čau, Vláďo:)
yop, the paper is exatly what I waited for. Thank you for that.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence