Week 6

 

My goals for this week are writing my DMP progress report, and continue using WEKA to see if I can come up with better results from my texture measures analysis.

 

Once I have implemented several different classifiers and ensembles, I will present my findings at the Tuesday meeting and ask for suggestions as to where to take the research

 

Though WEKA shows a high success rate, upwards of 80%, the confusion matrix shows that none of the crystals were classified as crystals.

 

At the Tuesday meeting I obtained the following suggestions: look at the images being sent to be analyzed, are they the right images? To add more crystal images, only one in 7 of my test images in a crystal. Add more attributes to the WEKA document. Check to make sure that all the functions are not misbehaving. Add a border to the image being analyzed. Try using a neural network classifier as opposed to the decision tree algorithms to make sure the error does not result from over fitting.

 

But before all of that goes on, I am to look at my existing data and compute descriptive statistics for the crystals and non-crystal images, namely the variance and mean for each of the 5 texture measures.

 

There does not appear to be any significant difference between the mean and variance of the crystals vs. non-crystals; in order to come to that conclusion.

 

I tried using a neural network, and that did improve my results but ever so minimally, so we can rule out over fitting.

 

A grad student, Pradeep, helped me import some images. After talking to a member of the vision group, I found out that the values I was working with (coordinates for the crystals) were actually shifted, so I had to take the coordinates I have and shift them before analyzing that area. Pradeep showed me some of the short cuts Matlab has for manipulating images and some unix simple codes to navigate the contents in several directories, and transport them to my working directory.

 

By Friday I have completed my progress report and submitted it, but my classification still does not produce any results. This is the first real he stumbling block hat I don’t seem to know how to get over. I simply have no idea why the texture measures are not informative, and since no one had ever tried using them to classify crystals, I had no guarantees that this would work. What if this is a dead-end? What if all my work over the past few weeks amounts to saying: ‘well, apparently it doesn’t work’? It is a scary thought.

 

This weekend I’m going to go see Pirates of the Caribbean with Johnny Depp. Also to a friend’s b-day bbq.