Archived

This topic is now archived and is closed to further replies.

Blazter

character recognition

Recommended Posts

How can I recognize numbers on an image? Does anyone knows a way to do or an article explaining it? I need help, it appears to be a little difficult to find information about this topic.

Share this post


Link to post
Share on other sites
The reason you can''t find much information is that there is no real solution for rocognizing numbers from any image.

There are however ways to do so if you have information about these images, or can have some assumptions. If the fonts are similar and the backgrounds the same, you can train a neural net to recognize it, or use some sort of template matching.

For more complicated cases you may have to do edge detection and further analysis.

Depends what you need this for, but I suggest you take a look at a computer vision book.

Share this post


Link to post
Share on other sites
I need to recognize numbers on car plates. The cars are positionated on the middle of the screen.
I''m studying how to do this, but for now it seems to me that''s it quite "difficult".

Share this post


Link to post
Share on other sites
Guest Anonymous Poster
Well if you had a dead on image of a few licence plates this might not seem such a cumbersome task. However you must realize that you will never get a 100% accurate answer.

A generic idea of how to do this would be to take the licence plate image. For the most part the top left corner of the licence plate could be used to determine the background color of the licence plate. You take this background color as your base color. Now seeing in real world images you don''t often have a "solid" color, you''ll need to use a ratio to determine if a color is in the same general range as the "plate color". You then go across the image pixel by pixel and look for a color which does not fall with in that range.

You can then evaluate the current colors which are left on the plate. Figure out which of these colors shows up the most in the image. Most likely this will be the color of the Letters/numbers of the licence plate. Use the range value to determine which colors are most likely letters/numbers and then get rid of anything that does not fall with in that range.

You should now be left with is some fairly ugly looking letters and numbers. You could easily use a bluring algo to go around the image and try to smooth things out in order for a more accurate match.

You then take this result, and compare it with actual computer generated letters and numbers. Use a ratio once again to determine what the error ratio you want to try is for each comparing letter/number before you say that this is not the letter/number you are looking for. Then you go letter by letter, and number by number comparing one image to the other. If the value falls inside of your error ratio you''ll save that letter/number out as a possability. Once you''ve finished this you can take the resulting letter/number possabilities and generate the possible licence plate numbers which this plate could be.

The more accurate your error ratios the less possabilities you''ll have. Now if the licence plate was skewed most likely you''d have to unskew the image, and run some type of guesstimation algo in order to fill in missing areas. but again this reduces the possability of a completely correct match.

Share this post


Link to post
Share on other sites