Week 7
Goals for this week are to find out why my texture measures are not working, see two dissertation defenses, research HSF graduate Fellowships, and attend a brain-storming for a Lego Mindstorms workshop for seventh graders that I have volunteer to lead.
In order to add more attributes to the WEKA document, I decided to compute the co-occurrence matrix using two different angles measures. That is, first process an area computing the co-occurrence matrix for 0 degrees, then process the same data but at 45 degrees, thereby effectively giving WEKA 10, as opposed to 5 attributes.
My mentor suggested to me that since good crystals usually
have sharp edges, not just single edge pixels, but small line segments, that I
should try to see if results from the Canny edge operator seem to correlate
highly with real crystals. In particular, threshold the Canny to get strong
edge pixels, take the connected components of those, and perhaps count how many
connected components of more than X pixels I get. I don't know what the proper
value would be for X. It has something to do with the minimum size of a
crystal. This is a simpler, less computationally expensive approach.
I would have to use Matlab to display the connected components of strong edge
pixels and see if they are helpful. The classifier ought to learn that such
components, when they are inside a textured area, may be just noise. But when
they are in a non-textured area, they are usually (not always) crystals. Air
bubbles sometimes give an incorrect classification as crystals. Previous work
by the vision group developed a procedure to remove them.
The problem is that you can get edges in texture. So my
mentor thinks I need to first find out if the area I’m looking at is textured
or not, because what I want to know is if there are edges that are NOT in
texture.
To that end, my mentor suggested to find several connected edges like a U-shape
and take the area of the convex hull as a feature. Again, it needs to be in a
place where there is little or no texture; otherwise it will falsely find edges
in the texture and think they are crystals.
I’m at a stage where I have moved from just
"implementer" to "researcher". My mentor thinks that at
this point, it is important to look at the sub-images I’m trying to classify
and really understand what is crystal and what is not, so that I will choose
suitable features. They don't have to be standard features. Here is where ideas
of my own would be welcome.
The "perfect" crystal looks like this:
/\
\/
How often does it actually show up like that in an edge image? What shows up
instead? When do you get a pattern like this or a subset of this that is not a
crystal but just some part of the precipitate or air bubble?
--
The meeting to brainstorm Lego Mindstorm workshop activities went quite well. The two grad students who usually lead these workshops will be unable to be present next Monday, so they have asked for volunteers to lead the session and I volunteer. I love to play with robots and I love the idea of bringing technology to the lives of kids. Julie, Benson and I came up with the activities, and the potential pitfalls we anticipated the students might have. Julie is to type the hand out we’ll give students, Benson is to procure the images and get the materials ready, I will have 3 assistants on the day of the presentation. I was complimented on my contributions to the team, thought it w as simple ‘You’re good’ without much ceremony, it meant a lotto me. I’ve had a wonderful experience this ummer, and I’ve felt welcome and that I am an integral part of the activities I’ve been fortunate enough to participate in.
--
At last weekend’s BBq, I talked to a student who is an NSF Graduate Fellowship Recipient, she suggested I apply for one such grant. I will need to come up with a research topic. I like computer vision, and my work with Linda has been very rewarding, I’m also interested in robotics and educational technology.