Feature Extraction takes a matrix of pixels, and returns a vector of information that is consistent for any one input matrix (the same input always gives the same output vector.)
The lower limit on the size of the vector for Feature Extraction is "empty", and the upper limit on the size of the vector is "what memory can hold".
The output vectors for two different input matrices are not required to have any obvious relationship to each other.
For example, one Feature of an input matrix could be obtained by reshaping the input pixels as a row vector and then using DES3 encryption on that row vector to produce the output row vector. There is no requirement that the output vector be short or simple to calculate -- just that it be deterministic.
If you have an algorithm that does this kind of deterministic transformation of a matrix to an output vector, then you have a Feature Extraction algorithm.
There are a huge number of different features that can be extracted from any input matrix. Feature Extraction doesn't care: if you supply even one bizarre algorithm then you are doing Feature Extraction.
That is what Feature Extraction is about.
Now, some features have been studied and are understood, and the studied features have different uses for different purposes. Some day you might wish to study some of the properties of those features. But for now, since all you need to do is "extract features", just make something up.
