Hello Prasad, for my short experience, I can tell you that:

For tecture LBP, Gabor and SIFT are a good starting point, and geometric (FAP) features, using Adaboost, mRMR and SVM feature selection algorithms could be helpfull..

Hoping it can help you..

According to my personal experience:

First I detect the interested region on a face, eyes and eyebrows, nose,  lip,

then I applied SIFT descriptor on every region independently. Thus my feature space was a set of key points from different region of the face (trains set). 

For testing you apply the same process on the new face image and compare with your train set.

You may take a look to these surveys:

A Survey on Facial Expression Analysis for. Emotion Recognition

Facial Expression Recognition: A Survey - ScienceDirect

Better LBP and its variants are better compared to others.

Thank you very much Girish G N,

                                for your kind response, if you don't mind i would like to take this opportunity to clear  few more  doubts. 

1) For creating a Data set of expressions, Which features are more suitable , say Contrast, homogeneity, etc., (or) is there any paper which best explains these?

2) I will be very happy if you provide me this LBP algorithm (step by step process) from giving image as input to getting feature vector as output 

Hi

You should try LBP which has its implementation in openCV. Also look into LDP with this paper http://onlinepresent.org/proceedings/vol17_2013/29.pdf. To use opencv go to this website http://opencv.org/downloads.html

Hi All

According to my experiance  normalized image face to fix size then divide image into mesh and the compute sift descriptor at coss points in the mesh. Apply LDA or PCA transorms to increase the descmenative power of features vector

Dear colleague,

I can suggest you to read this article:

http://visgraph.cs.ust.hk/biometrics/Papers/Face/pr2001-10-2.pdf

I am pretty sure it can help.

Kind Regards