search

MatInfTeam4 / MatInfTeam4.github.io

The domain website repository for Materials Informatics class project, Fall 2014.
MatInfTeam4.github.io
MIT License
0 stars 4 forks source link

Explore Segmentation Methods to Extract Grain Boundaries #5

Open ahmetcecen opened 9 years ago

ahmetcecen commented 9 years ago

Especially at highly deformed/ambiguous regions.

pfz3 commented 9 years ago

I got 100X-mag pictures in the dropbox link. The images look pretty cruddy. I played a little with the edge function in matlab and tried different filters. I can take SEM images if it will alleviate the situation.

pfz3 commented 9 years ago

I played a little more today with multi-thresh techniques. Getting rid of low luminosity pixels gets rid of the epoxy regions and helps focus on the metallic part. This doesn't work on the interior regions, but if we handle a "stiched together" image this won't be a problem. I think we can discriminate the shear bands using a line-filtering type of technique as we discussed in class. This will help us identify "segments" in a much more automated way.

Oh and I uploaded the matlab script I used to the images/22 folder. Right now it looks for the images in my local dropbox folder. I'll work on getting that to pull from a url.

ahmetcecen commented 9 years ago

I have had more promising results with some alternate techniques. I will try and create a summary post by Wednesday that goes over every segmentation method we have tried on the initial dataset. I believe we might be able to use the data we have to some extent.

pfz3 commented 9 years ago

I played around today and I noticed something funny about all the images. In each image, if you zoom in on the middle of a rather large grain, you can see a repeating pattern of small circles. It appears as this pattern occurs everywhere in the image. I checked on the microscope and found that when I move the stage, the small circles don't move. I believe the lense is dirty. I read a bit on the edge detectors / gradient calculators and I believe that these small circles are introducing alot of error in the gradient calculations and hence the edge detectors fail... I couldn't find a way to take derivatives over a larger range of pixels, so instead I reduced the resolution of the image. This appeared to "collapse" the small circles and the edge detector works VERY well. I'll add a post or something soon.

ahmetcecen commented 9 years ago

Yes, Tony and I noticed the same thing on Monday when we were looking through the ten slices that contribute to a multi focus image. Reducing the resolution does help a lot, as well as using the three channels of the RGB image separately for segmentation.

pfz3 commented 9 years ago

Ok nice. Hadn't thought of the RGB thing. Are those dust "specs" mostly comprised of R or G or B??

pfz3 commented 9 years ago

Hey Ahmet someone from my lab is leaving and we are having their good-bye lunch today. I am going to lunch at noon. I may be back as late as 130/2. My phone is still dead so you will have no way to contact me... I think that the post I wrote for today's class may be bare-bones enough for today's presentation but if you want to add stuff or refine it feel free. I'll check my email when I'm back from lunch.