Help with Super-Resolution work on a license plate please

[berkeman]

I have several images of a California license plate (blue letters on white background) that were taken by a security camera here at my work. We need to contact the driver of the vehicle, so I am trying to apply some processing to the images to see if I can make out the license plate number.

I did some googling, and found that the process of combining multiple noisy images to get better clarity is called "Super Resolution Reconstruction". Does anybody here on PF have experience with this kind of image processing? Is there a different approach that I could also try? I downloaded a free image processing program called ALE, which is supposed to include SR capability, but I haven't had time to figure it out yet.

There are 5 base images of the plate, and one PhotoShop enhanced image of the best plate image. I've attached three images to this post, and I'll post the other three in the next post. Thanks for any help or ideas you folks have.

 

[berkeman]

...and here are the last 3 images:

 

[berkeman]

BTW, for privacy reasons, it's probably best not to post guesses at the plate or processed images in this thread. Go ahead and talk about the issues and that you think you have a good guess or image, and I'll PM you my e-mail address for further contact.

Thanks again for your ideas and help.

 

[AO eye 5]

I have not look at the images, but what you are describing seems a lot like speckle interferometry.

Edit: Just my observation if it helps.

 

[berkeman]

BTW, an initial guess with the PhotoShopped image was 44M2O57, but that turns out to be a non-valid plate. It doesn't follow the convention for California standard plates NLLLNNN (N=number, L=letter), so it would have to be a custom license plate. But that custom plate hasn't been taken yet, so it's wrong.

 

[J77]

Are those attached bmp's the best you've got?!?

 

[berkeman]

Are those attached bmp's the best you've got?!?

Yeah, pretty stupid, eh? A good camera would have that driver, um, contacted by now. Unfortunately, it's up to signal processing to make that contact now...

 

[Evo]

It always kills me that companies invest in security cameras then go dirt cheap on the cameras which are a one time cost, the result being that the entire system becomes worthless for identification.

 

[NoTime]

Why does the plate look yellow and red?
You said CA is white and blue.
You sure it's CA and not out of state?
An out of wack camera to make white yellow would most likely turn blue to black, not red.

 

[berkeman]

Why does the plate look yellow and red?

The pictures were taken at night of the front license plate. The ambient lighting at that time was from yellow-tint street lights. I should have mentioned that in my posts.

 

[NoTime]

Ummm.
Yelow light would be yellow because it is mostly red and green wavelengths and low on blue.

Yes, that would turn white yellow.

However, the blue should still show as black.
Because there is no blue for the paint to reflect.

What appears to be red lettering would be inconsistant.

Given your environmental conditions, I would go for a white plate with red lettering.

 

[NoTime]

Out of curiosity I poked around for a few miniuts.

I'm going to guess it's a New Mexico plate.

 

[DaveC426913]

44M2O57?

I just saw a car with that plate. I keyed it. I hope it was the right thing to do.

 

[Evo]

44M2O57?

I just saw a car with that plate. I keyed it. I hope it was the right thing to do.

You're bad.

 

[chroot]

I suggest looking into track-and-stack image processing programs. These are commonly used for tasks like webcam astronomy, which produces a succession of low-resolution images. You can align and add the images (hence the name track and stack) to improve SNR and perhaps see more detail that on any single image alone.

You can do the same work with a tool like Photoshop, as well -- blow the images up, artificially increasing their resolution, then align and add them together.

- Warren

 

[berkeman]

Thanks Warren. Track and Stack gave me some good hits on google. I'll let you know what happens.

 

[chroot]

I'm trying to do some track-and-stack like work in Photoshop, but it's not revealing much. I really think the SNR of all but the first couple of images is so bad that they're not worth using. Unfortunately, you only have a couple of the better-SNR images. If you had perhaps 10 or so frames of that quality, this would probably be tractable. Do you have any additional images at all?

- Warren

 

[berkeman]

Unfortunately no. The security camera presently only shoots 2 frames per second, and has mediocre resolution. Hopefully that will be upgraded in the near future. Thanks for trying, Warren. -Mike-

 

[franznietzsche]

Out of curiosity I poked around for a few miniuts.

I'm going to guess it's a New Mexico plate.

Thats what I thought too when i first saw it, but then the letters would be orange.

Incidently, its worth keeping in mind it might be a 6 character plate, especially if it ends up not being a california plate.

 

[NoTime]

Thats what I thought too when i first saw it, but then the letters would be orange.

Incidently, its worth keeping in mind it might be a 6 character plate, especially if it ends up not being a california plate.

I was thinking the "Balloon" plate.
This from pictures has a hot air balloon on left and red lettering. Not the older? plates.

If you mask out the red and green channels it vaguly look like a circle on the left side.

 

[mmccorma]

Super-Resolution License Plate

I applied super-resolution techniques to your images but the result was not good. Probably they don't contain enough information when combined to produce a higher resolution image. See attached image for an example of 5 low-res license plate images (LR) and producing a good super-resolution (SR) result.

--Mike

 

[berkeman]

Thanks for tryin, mmccorma, and welcome to the PF. What software did you use?

 

[mmccorma]

No problem. I wrote the software and a GUI for super-resolution in Matlab. It finds the integer and then sub-pixel offsets among a set of images using the Taylor series method described in this paper: http://www.ph.tn.tudelft.nl/~lucas/publications/2000/SPIE2000CLLV/SPIE2000CLLV.pdf

Finding the correct sub-pixel offsets is the key to resolving a higher resolution image. Then it's just a matter of interpolating a non-uniform grid of pixels to a uniform and denser grid.

--Mike

 

[DaveC426913]

All you need to do is send your picture to the CSI lab in Vegas for analysis.

They have magical software that can take a blurry, low-rez image of a whole bank interior, zoom in on a credit card in someone's hand, pick out the individual numbers on it, and display them in perfect clarity. Now that's enhancement!

 

[mmccorma]

Ok, here you can see how CSI does it :) ie. zoom-in on the license plate movie: If the link doesn't work just search for Super-Resolution.

 

[ray b]

from mike's post I read _D267
with the _ looking like the handycapped tag wheelchair grafic