Ask Your Question
0

Number plate LBP training curiosity

asked 2017-05-29 21:06:41 -0500

Hi,

I've been reading a lot lately in order to train OpenCV to find number Australian number plates in an image. I was just wondering : from what I understand of the training process, you provide a lot of real life images of the object you want to train your detector for (I have access to thousands of images that's fine) in different conditions..

Now, a number plate is, by essence, unique. So just out of curiosity here (because I know some LBP detector do work for ANPR) : is that why the detector need soooo many examples ? Would not it be "easier" to train it to find objects looking like number plates : find the rectangular shape but not caring about "what's inside" ?

Maybe I got it all wrong; I just woke up with this question in my head seeing my training having run for 16hours and still in phase 1 ^^

Cheers guys.

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
0

answered 2017-05-31 04:29:21 -0500

Let me reply based on your comments

Now, a number plate is, by essence, unique.

That is true, but by providing many samples, you let the boosting look for features that are general to the class, and not specific to one plate.

So just out of curiosity here (because I know some LBP detector do work for ANPR) : is that why the detector need soooo many examples ?

Yes because basically it wants to generalize. To have a clue on what it does, take 1000 images of numberplates and convert them to their LBP counterpart. Then normalize/average over the 1000 plates. The resulting average plate will have the features used in your learned model.

Would not it be "easier" to train it to find objects looking like number plates : find the rectangular shape but not caring about "what's inside" ?

No, because you are tackling it visually. Features are capable of capturing relations that we do not see with our eyes. If we just use rectangular shapes, you will have way more false positives to get rid off somehow.

edit flag offensive delete link more
Login/Signup to Answer

Question Tools

1 follower

Stats

Asked: 2017-05-29 21:06:41 -0500

Seen: 259 times

Last updated: May 31 '17