0
$\begingroup$

My first question - This might be a basic question but I have yet to find an answer; when choosing the features for my model, I have encountered certain features which are vectors themselves. (e.g. Haralick texture Feature, Gabor features, HoG, etc.. as you can probably tell, I'm working on an Image classification problem.). Yet, in the literature I've only found reference to cases where each feature is a single number.

So should my feature vector look like so : [[a,b,c],[d,e,f],g] of like [a,b,c,d,e,f,g] where I 'ignore' the brackets? if so what are the implications?

My second question is - I'm trying to design a classifier to detect whether an object is a cell or not from photos obtained by microscope. but my problem is the cells I'm working on differ from each other by big margin, which creates a big problem when trying to distinguish them from all other objects seen in the image (sand, dirt etc..). How should I attack this issue? Are there known approaches?

Regards.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

To address your first question, if [a,b,c] are treated as three independent features, then the algorithm may select, say, a but not b or c. This might mean using one third of a Gabor filter or half a histogram of gradients etc. which seems a little strange and hard to interpret. However, presumably that would only happen if the feature selection algorithm discovered that the independent components are useful by themselves, so flattening the vector (ignoring the brackets) makes sense.

For your second question, if there are several cell types that differ widely, it may be simpler to build multiple classifiers, one for each type. Alternatively, you could develop classifiers that identify and exclude the negative objects (sand, dirt etc.) leaving just the cells.

Improve this answer
$\endgroup$
2
  • $\begingroup$ thanks for your answer. For the second question - wouldn't it be more computationally expensive to develop classifiers for each type? even so how would that address the problem of 'negative' samples? what I've started doing is designing a classifier that first decides if the given object is a cell or is not. if it is a cell then use a different classifier to decide what type of cell. but the first problem proved to be very hard. for your second suggestion - i did not understand what you meant by "exclude the negative objects leaving just the cells"? how would you go about doing that? $\endgroup$
    – user80280
    Commented Jul 20, 2015 at 13:27
  • $\begingroup$ Developing multiple classifiers will require more computation, but may still be the best solution. It's hard to imagine, but if you consider the (high dimensional) feature space, it may be that one region contains one target class, and another distinct region contains instances of a second class. Many classifiers are best at finding such distinct, convex regions, so multiple independent classifiers may be appropriate. By "excluding negative objects" I meant training a classifier specifically to identify (e.g.) dirt. Then the main cell classifiers don't need to even consider those objects. $\endgroup$
    – dcorney
    Commented Jul 21, 2015 at 10:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.