How to extract License Plate?

I want to extract license plate area but couldn't figure out how to remove extra region or directly extract the license plate area.

[Image Analyst]

It looks like you've already identified regions because you've blacked out some bands. What information do you have? Do you have starting and stopping row and column numbers?

[Cedric]

Could you post a series of full pictures? There is no way to write a generic algorithm, but if we have original images, maybe there is a simple set of tricks which could work specifically for you case.

[NS]

@image analyst. sorry I didnt notice your message. I scanned the image horizontally and vertically and set the certain threshold of every image by calculating the difference between pixel. Then I blanked the region which have low threshold than average threshold. And regions which have high threshold than average threshold are shown above(in high threshold region there is always a region of license plate). Now I want to extract the license plate region. I dont have starting and stopping row and column numbers.

[NS]

@Cedric Wannaz

Original Image above image: http://s2.postimg.org/lrvj0scwp/Image297.jpg

Original image: http://s16.postimg.org/9bxsmxko5/Image397.jpg

Final image i got so far: http://s15.postimg.org/i7vcgg0az/untitled.jpg

[Walter Roberson]

Are the license plates to be recognized always white letters on a black background? If so that is important information that allows you to throw out the other sections of the image you have identified, and would allow you to discard the tricky Japanese example I posted as being Out Of Scope. It would still retrain the Afghanistan plate example, though, as it is white letters on a black background.

[Cedric]

So all images will be similar in the sense that they are taken by hand at a distance of a few meters and roughly centered on the license plate? And all license plates have a black background and white symbols made of three large letters in the top row plus an optional extra with a smaller font, and between tow and four centered numbers in the bottom row?

[NS]

white letters on a black background and three large letters in the top row plus an optional extra with a smaller font, and between tow and four centered numbers in the bottom row is not always true. Some plates are opposite to this. Link this. http://s3.postimg.org/fbw3b7dhv/Image27.jpg

http://s2.postimg.org/fwb8o8tax/Image76.jpg

http://s8.postimg.org/6nr4c0dv9/Image119.jpg

1: distance between camera and plate is always fixed.

2: license plate is approximately in the middle of image.

[Walter Roberson]

Cedric, distinguishing letters and numbers requires an OCR step. The poster... "requires"... something that does not do OCR at all in order to select the region to submit to the OCR stage. We don't know -- perhaps the poster is being charged $50 for every section the poster submits to OCR. T'is not for us to wonder why.

[Cedric]

Walter, we can obviously not provide a full solution (stable, solid, etc). I was asking because if the OP is stuck not being able to extract the area where the license plate is before submitting to OCR, it may be possible to find a simple trick (without guaranteeing performance/stability, a bit like what I proposed here), for example based on a simple convolution, which succeeds in e.g. 60% of cases and is able to flag the 40% remaining as failures that must be processed by hand.

[Cedric]

Nomi, seeing now the large variability among license plates, fonts, colors, geometries, distances from car, poor centering, I don't see any simple trick that would apply to all cases.

Answer 1

[Walter Roberson]

http://uk.mathworks.com/matlabcentral/answers/?term=tag%3A%22lpr%22

[NS]

First thanks for the above link. It really helped me. But still I cant figure out how to differentiate between text area with non text area. (I thought this might be the solution but not sure)

[NS]

I have read somewhere that Euler Number is helpful in differentiate between text area with non text area. But i don't know how to implement it.

[Walter Roberson]

http://uk.mathworks.com/matlabcentral/answers/34679-please-can-anyone-help-i-require-a-matlab-code-for-removing-text-from-an-image

[NS]

I can't apply OCR at this stage. It detect much more characters than on license plates (and obviously detect wrong characters too). I think OCR works well on correctly segment plate.

[Walter Roberson]

If you had read my Answer there you would see that everything is text. A picture of a bird is text -- though you might not understand the language.

[NS]

I got your point. But I just want to detect the License plate area. Any way to do it?

[Walter Roberson]

You are still asking the wrong question.

[NS]

I didnt get your point. Can you provide me its coding or any algorithm?

[Walter Roberson]

Everything about images is interpretation. There is no rule that can reliably distinguish license plate from non-license-plate or text from non-text.

The most you can do is to restrict yourself to defining a particular kind of license plates and a particular kind of text that you are willing to restrict your project to. http://porcelainplates.net/gallery_shapes.html

[NS]

I cant restrict myself to certain type of license plates. Here We have very variation in terms of license plate. Some are small, some are large, some are long, single line, double line etc.

That is why I calculate the threshold of every image and shows the regions have higher threshold value than average value.

[NS]

Now I am trying to calculate the gradient of every region and then shows the region with highest gradient value. Because I guess license plates have always highest gradient value. Hope this works.

[Walter Roberson]

If you can't restrict yourself to certain type of license plates then you are back to the problem that anything can be a license plate, since "license plate" is an intention. It would, for example, be perfectly valid for a jurisdiction to use the 143000+ different paintings of Morris Katz as license plates.

Answer 2

[Image Analyst]

See this: Automatically Detect and Recognize Text in Natural Images

[NS]

I have tried it but not getting right results. Dont know why.

[Image Analyst]

This is not a trivial task. If the built in capability of the Computer Vision System Toolbox is not working for you then you'll need a much more sophisticated algorithm. It's not some 5 minute job that Walter or I could do for you. People have written Ph.D.s on it. Luckily they did the hard part - inventing the algorithm - so all you need to do is find the one, from the dozens that are out there, that works best with your images. Presumably you're not planning on being a LPR algorithm developer yourself, and just need it, like a black box, to continue on with the main purpose of your project. By the way, what is that? If you had the license plate number, what is the main thing your research would do?

[NS]

Yes I know its not simple. I struck at this point. I cant figured out how to get rid of this problem. Hence just asking for the right technique.

[Walter Roberson]

The right technique is to define the problem to one that can be solved. Like only having to recognize 1910 license plates from the US state of Maryland.

Answer 3

[muayad bakhtan]

there are an efficient algorithm that had ability to extract the region such as Edge detection and histogram and color map .. you can back to one of this algorithm .

[Walter Roberson]

Refer to the original image that the user is working with.

You can see that this has numerous edges, but it does not have a clear edge around the license plate. You also cannot use color of the plate background to tell, as the background of the plate is the nearly the same black as the car itself. Color histogram would probably tend to suggest that the color to look at would be silver rather than white, under the assumption that the most prominent color would be the car body and that the second most prominent color would be the car lights and that the plate letters would be the third most prominent color. On the other hand you have to take into account that lights are usually two colors together... now is the white of the lights more prominent than the silver or the white of the license plate? Don't forget to account for the fact that in many cars the bumper is a different color than the car body.

Could we arbitrarily focus on white? Sure, but the poster said specifically that the algorithm must work on a variety of plates of different shapes and colors and colors of letters.

You should have another look at some sample license plates I posted earlier in http://uk.mathworks.com/matlabcentral/answers/232793#comment_303089

There are no efficient algorithms that can handle all license plates. There are no inefficient algorithms that can handle all license plates. As I wrote earlier, "license plate" is an intention, not a particular size or shape or color or font.