i found out that HOG features are best when compared with eigen features or fisher features. i don't know particularly about the spine images in specific, but i can tell you that you can start with finding the HOG features. 

i worked on face images. you can see my paper explaining about finding the HOG features and the importance of HOG features. 

instead of spending time in identifying the methodology for the best feature vector, you can work with already existing and experimentally proven features one by one. i hope you will be succeeded with the first experiment.

MATLAB is the best platform for finding the features for images. 

plenty of MATLAB programs are available in the internet.

previously i tried other features like SIFT and so on. i didnot get as much output as HOG features. 

all the best

You can compute Fisher Ratio=(within Vectors Variability)/(Between Vector Variability) or other statistical analysis technique.

Assuming that you use the feature vectors for a classification problem, why don't you use the features on your dataset and experimentally determine which one is the best? The classification results should be good indicators. You can use a one-leave-out cross-validation scheme to determine the robustness of your feature vectors.

Other than that, you can try to calculate mutual information between the feature vectors and class labels. I have done that in one of my publications linked below. 

Conference Paper Ensembling brain regions for brain decoding

Thank you Alkan.  I will certainly go through your publication. I was just wondering how to decide what is the best feature vector for a certain problem without having to jump into the classification stage. How to define the variability amongst feature vectors?

Jiyo

Basically this is not easy to answer which feature is the best. Alkan has truth answer but actually a correct way is that why you selected feature X or feature Y.

If you can answer what was your aim to choose an specific feature, then you can say which one is the best or must be the best for your specific goal. 

Briefly, you have to know what is the properties of the features, for example HOG is a feature that extract gradient information, SIFT is scale invariant, LBP works on intensity, etc. So you have to see what you need otherwise it does not make sense to use features randomly.

best

i found out that HOG features are best when compared with eigen features or fisher features. i don't know particularly about the spine images in specific, but i can tell you that you can start with finding the HOG features. 

i worked on face images. you can see my paper explaining about finding the HOG features and the importance of HOG features. 

instead of spending time in identifying the methodology for the best feature vector, you can work with already existing and experimentally proven features one by one. i hope you will be succeeded with the first experiment.

MATLAB is the best platform for finding the features for images. 

plenty of MATLAB programs are available in the internet.

previously i tried other features like SIFT and so on. i didnot get as much output as HOG features. 

all the best

Dear Mahdi, Thank you for the response. Yes, I can now get a picture of which feature to use. I actually had few in mind based on my context, but just wanted to test them on a similar ground.

As for HOG,eigen and fisher features, Mr.Harihara, thank you for the suggestion. I'll certainly try them if my problem needs it. 

If features you extract provides complementary information then you can try features fusion and do boosting to find most discriminant features.  

You can use one of several distances which exist in the litterature, like Euclidian distance, or other...

but the problem depends on your dataset experiment. the good way to compare or select the better features is to compare the résults of your experimentation.

The easiest way is to Euclidian distance, between the test feature vector and training feature vector. Select the class which giving less Euclidian distance is the best match.