Before you go too far down the road using Gabor features, you might do well to read recent results achieved using deep learning:

http://fastml.com/how-to-get-predictions-from-pylearn2/

http://kaggle.a.ssl.fastly.net/c/dogs-vs-cats/forums/t/6845/example-use-decaf-via-nolearn-for-94-accuracy/37698

The state of the art on image understanding has advanced dramatically in the recent few years.

I am not sure which method is the best. However, ANOVA is a good start helping you to select the most significant features. In addition if you are going to classify your images or objects, based on the features, linguistic hedge variables are powerful tools in neuro-fuzzy to select the discriminative features:

http://www.sciencedirect.com/science/article/pii/S0957417410001545

thanks!

I found this article very helpful for my study and maybe it can help you too :

http://bioinformatics.oxfordjournals.org/content/23/19/2507.long

It is a review of feature selection techniques discussing the advantages and drawbacks of each.

Good luck !

The article by Saeys appears to be comprehensive with regard to feature selection, but in image analysis feature selection is commonly not used.  Regularization is used quite commonly instead.  In the case of deep learning, various means of regularization are combined with learning to extract the features in question.

As such, I think that focusing on feature selection may be a big mistake.  The example I gave involves no explicit feature selection at all and yet it gives very good performance on a quite difficult image analysis problem (cat versus dog recognition).

I think feature selection is mandatory in image classification because we don't know which feature descriptor is good to represent the images.

Hello, Arun! I'm not able to tell you which is the best algorithm for feature selection in order to solve your problem. However, I suggest you to research about the Successive Projections Algorithm. It is a very helpful technique for variable selection in many type of problems. For instance, I have used this algorithm for variable selection in multivariate calibration problems. Perhaps it may be useful for you too. Please, take a look at these two papers of mine.

Conference Paper Partial Parallelization of the Successive Projections Algori...

Article Parallelization of a Modified Firefly Algorithm using GPU fo...