Re: Feature Extraction - Range breakdown

Hi,

GIFT does not do any feature selection. Many features are binary and =

thus either present or absent.
The different number results from some images containing for example =

fewer colors than other images and from the fact that not all images =

have all direction/scale/strength combinations of the Gabor filter =

responses.
For the Gabor filters that default configuration does not use the lowest =

energy band, meaning the images with many flat regions have fewer =

features. This is an option in the source code and can be changed.

There is a feature description file in the index directory that has the =

feature number and feature group (on of the four Gabor histogram, Gabor =

block, color histogram, color blocks). I do not remember the exact =

details but it should be described in the early papers on the system =

such as: http://publications.hevs.ch/index.php/publications/show/189

The Gabor blocks are indeed 16x16 blocks, where as the color blocks =

should be blocks at four scales, so first dividing the image into 4, =

then 16, 64, 256 blocks, so 340 blocks if I am not wrong.

Kind regards, Henning

On 12/19/11 12:44 PM, Xavier Clements wrote:
> Greetings,
>
> I have two questions, the first having two parts.
>
> 1. How does GIFT select features for extraction given that I read in one
> of these discussions that the range of features extracted for any one
> image is between 341 to 3686? Is this upper limit arbitrary in nature
> and able to be altered in the code?
>
> 2. Is the following breakdown of the ranges for each of the four
> extracted feature types correct (taken from a previous post)?
>
> "1. color blocks
>
>
> - 1st 256 times 16x16 color blocks, where each has 166 possible values
> (i.e. features 0..42 495, where 0..165 are from the first block etc.)
>
> - 2nd 64  32x32 blocks,
>   (10 624 more feats, i.e.42 496 .. 53 119  <tel:42%20496%20..%2053%20119=
>)
>
> - 3rd 16  64x64 blocks,
>   (2 656 more, i.e.53 120 .. 55 775  <tel:53%20120%20..%2055%20775>)
>
> - 4th 4  128x128 blocks
>   (664 more, i.e. 55 776 .. 56 439)
>
> 2. color histogram
> 166 features
> (feats 56 440 .. 56 606)
>
> 3. gabor blocks
> - 256 times 16x16 blocks, each block containing 120 possible Gabor featur=
es
> (features 56 607 .. 87 325 so that in 56 607..56 726 are the Gabors calcu=
lated
>
> from the first block and in 56 726..56 845 <tel:56%20726..56%20845> are
> from the second etc.)
>
> 4. gabor histogram
> last 120 features ( 87 326 .. 87445 )"
>
>
> Forgive this newcomer for potentially asking questions that have been
> previously answered (I looked through a number of the forum's
> archives...honest).
>
> Thanks in advance,
>
> Xavier
>
>
> _______________________________________________
> help-GIFT mailing list
> [email protected]
> https://lists.gnu.org/mailman/listinfo/help-gift

-- =

Henning M=FCller, PhD, Prof.
HES-SO Valais
TechnoArk 3, 3960 Sierre, Switzerland
Tel. +41 27 606 9036
http://medgift.hevs.ch/