Down sampling

heyy guys this is Riti in image processing m having large set of feature vectors now(after applying different types of filtering technique), now i have to reduce the features using down sampling.. which type to down sampling should i use and how to use???

Answer 1

[Amith Kamath]

I would assume that in this case, a more apt term is dimensionality reduction, which has to be done based on what features have the most distinguishing power, and that, in turn has to be found out based on which learning model you are using. I know this answer seems very generic, but so is the question! You may find PRtools linked here quite interesting:

http://www.prtools.org/book.html#ProjectS

Thanks!

[Riti kushwaha]

thank you amith... thanx for this tool..

Answer 2

[Image Analyst]

I assume, like Amith, that you don't really mean down sampling. So you might look into principle components analysis, and the like. If you have a statistician available, I'd definitely talk to him or her for assistance. They can help a lot.

[Riti kushwaha]

actually currently i m having a huge feature vector. approximatly 40(orientations)*(137*128)(image size) features.most of the feature are retundant.so firstly i have remove that redundant feature. the whole process of removing redundant feature, is called(by me) down sampling. then after i will apply PCA LDA and all...

thanx for the reply...

[Walter Roberson]

"down sampling" has a very different meaning than that in image processing and signal processing, so that term is getting in the way of people understanding what you are trying to do.

If you have redundant features, then why bother to calculate those?

If you do not know ahead of time which features are redundant, then what will you use to determine afterwards which features are redundant? Will you, for example, be looking for duplicate rows (or columns as appropriate) to see that one feature had all the same outcomes as a different feature?

[Image Analyst]

You have "40(orientations)*(137*128)(image size) features."??? So it appears that the number of features you measure depends on the number of pixels??? This is very unusual. Usually you might have only a handful to a few dozen features, not thousands. By the way, the pixel value itself is not a feature. Can you tell me why the number of features that you measured depends on the number of pixels?