Color/texture matching
Daniele Vicario from Italy  [1 posts]
10 year
We need to quality control several different granite colors (reproduced by paint).
You can see two examples in the attached pictures.

Can Roborealm be used for that purpose ?

Basically it should return the match of the inspected sample compared to the reference color.

Each granite color is composed by a background color plus dots of one or more (up to four) different colors.

The system should check:
- background color matching
- dot's color matching
- dimension of the dots
- spacing of the dots

Thanks for any ideas.


Juan Cabrer from United States  [11 posts] 10 year
Given the number of variations that are possible, this type of inspection needs to be able to "learn" wht is good and what is out of spec.  That is assuming that the inspection is to be automatic.

It's not a simple thing, but RoboRealm is definitely capable.  You will need to identify the variables that are important to the inspection, and capture them during a teach step, where hundreds of good samples are taken in for each "product".

Once you have these samples saved off, there has to be a way to identify the product.  This can be done manually, or there can be another RR function to pick the closest match.
Anonymous 10 year
Thanks for the hint.

I didn't think of a learning approach and I'm not sure it's going to be usefull in my case, or at least I don't know how to implement it the right way.

I tought of a more "classical" approach.

The basic idea is (or was) to identify each dots and classified/group them by similar colors. I think this could be done using blob functions.

Counting them bring the average density of each colors (since the picture scale is fixed) while analysing each group can take to the min/max dimension of the dot.

Once you obtain these information you'll end with some vectors/numbers (background color, densities , colors and dimensions of the dots) that should be classified.

After that result I think you are right, you need a sort of training algorithm but... well any help it's really appreciated.

Steven Gentner from United States  [1446 posts] 10 year

You'll have to have a couple images of when a match is NOT considered correct. There is a bit more research that is needed here. The idea being that you would process the original image in some way, generate some sort of meaningful statistic and then compare that statistic to the 'gold' standard. This will differ for each pattern.

For example, if you look at the attached it isolates the spots from your first image which can then be analyzed in some way. In this case, the average blob size can be calculated ... but that's meaningless unless we know that a blob size of < 10 and > 60 is good. That's what your good versus bad image tests will tell you.

The research is then what quaility dictates what is good verus bad. Average blob size is a simple statistic but it might not be sufficient. You mention density (total blob area / image area) which can also be calculated but without comparing that to good versus bad images you'll not know what to do.

Thus, start by building up a database of good and bad images and start looking into what statistic is the most obvious one that differentiates the two. The blob filter will be helpful in that regard.

Its not a simple problem so you may need to reach out to someone with the math or analysis skills in order to complete this. A training algorithm can help but you still need to know what it would be training on (i.e. blob size, blob density, etc.)



This forum thread has been closed due to inactivity (more than 4 months) or number of replies (more than 50 messages). Please start a New Post and enter a new forum thread with the appropriate title.

 New Post   Forum Index