Defect identification and classification

[zulu]

Hi

Can anyone direct me the possibilities in labVIEW IMAQ Vision to find and classify the defects on the piece of sample attached.

[Brian_Beal]

This is pretty easy. Some more defect samples would be needed, and maybe some different lighting, but as it is now, this is no problem.

[RManion]

Hello again,

I made some additional comments on the other copy of this thread. For the sake of clarity, if you should have any other questions or information, please pick one of the threads to post to. This way, we can be sure that all of your questions are answered.

I thought that I would also note that it looks like Brian has used either a threshold or some form of edge detection. Try these functions out and let us know how it goes!

Thanks,

Robert

[taufiqhabib]

This is a classic problem in web inspection.
I would ideally recommend a texture based operation.
With the IMAQ Vision library, you should use a convolution and/or a morphology function as the first processing step. You want to be able to use the geoemtrical and frequency characteristics of your defect.

The problem with using straightforward thresholding is that this method is sensitive to the eveness of lighting as well as the stability of light output. Also you will find yourself continually adjusting the threshold.
Even with the smarter auto and local thresholding methods, you not find this method to be robust.

Also be aware that you should try to accumulate as many defect samples as possible to account for variations in the background and the defect itself.
You may start this machine vision project to find and fix a certain defect you are seeing now. Once you control and/or quarantine this defect, you may find other defects you want to detect.

In order of importance, your vision solution should be:

1. Robust - not require frequent tweaking of parameters due to variations in the product (fabric), process, and defects.

2. Fast - to keep up with your process throughput.

3. Flexible - to allow for changes in the detection algorithm

Message Edited by taufiqhabib on 03-17-2005 12:09 AM

[zulu]

Hi Taufiq,

Am able to find out some coarse defects which is visible to eyes.
i need a more robust algorithm to identify and classify the kind of defect.The defects can be torn pieces,shade variation,invisible foreign matters ....
Is there any possibility to achieve this.
Thanks for the threads fwded and i found out that cooccurence matrix and its parameters can differientiate textures i.e type of cloth.
I wonder will cooccurence matrix and parameters can classify the defect.

eagerly awaiting
zulu

[chilly_charly]

---

@zulu wrote:
I wonder will cooccurence matrix and parameters can classify the defect.

---

One way of improving the detection of a texture variation (ie the presence of a defect), is to calculate the texture parms on a moving window of reduced size instead of the whole cloth image. That way, you will be able not only to classify but also to locate the defect.
Consider applying statistical methods such as PCA (Principal Component Analysis) to find the best combination of parameters that maximize the "distance" between defects.
Remember also that there are 18 other texture parms that were defined by Haralick !

CC

[taufiqhabib]

Zulu,
I'd like to pose some questions for you here...

Texture analysis is ideal for this job but be aware that this is computatinally intensive. How fast is the fabric moving (if it is)? Web inspection requires very fast vision processors or many of them in parallel. So what is your budget for this project ?
Also do not underestimate the complexity of this task .. as with most vision projects you need to manage the expectations for your customers and managers. So take the time to do a formal study and do some research on web inspection. This may take a few weeks or more depending on your current vision experience.

Use some of the modules provided by peole here on texture and get your 'hands dirty' doing inspections on a test bench setup. Try some other faster non-texture methods.
Adaptive thresholding followed by binarization is the fastest and may be suitable if your lighting is well controlled and the fabric is moving smoothly.
Filtering and morphology lie somewhere in between texture analysis and binarization.

And by all means quanfity and classify your defects.
For example:
Dark vertical defect
Light vertical defect
Dark horizontal defect
Light horizonal defect
Dark spot defect
Light spot defect

Hope this helps!
Taufiq

[taufiqhabib]

Charly,
I agree that the moving window method is much faster then doing texture on the entire field or view.
The window can be appropiately placed once a 'coarser method' indicates a defect may be present.
One would also need to account for the possibility of multiple windows in a a single FOV

Taufiq

[abbas]

Hi Charly,
Your proposed method is quite suitable for my work, could you please explain how can I use moving window to locate the defects? Is there a code in Labview to do this?
Thanks a lot
Abbas

[chilly_charly]

Unfortunately, there is no example code published on the subject. You'll have to develop your own code to do that.

Bon courage !

CC