Geometric features are sensitive to expressions while PCA is sensitive to illumination variations. Today's research focus is on developing algorithms which can handle variations in pose, illumination, expression and occlusion. Techniques such as Discrete Cosine Transform is said to have good information packing ability and can capture features which are less sensitive to the above variations. You can also use Wavelets as it is possible to apply multiresolution techniques on face images to capture more valuable features. Further your classifier must be strong enough to recognize faces correctly.

In general you can use AAM or ASMs to locate fiduciary points on the face. These points can the be used to extract features such as eucledian distance etc. to classify expressions/ faces.

Hi, Basic concept you can clear with these link below. I already work on face dataset for using PCA. 

basic steps 

Read your image(m X n)

Resize image into a vector (1Xk) where k=m X n

Assume you have 10 image for training purpose then size will 10*k

Now find covariance matrix of 10*k image dataset

Furthermore, find Eigen-vector of same dataset

Now select first top m feature from dataset so, your final matrix will 10*m.

Now you need to do same procedure for testing image and compare with eigen face dataset using some distance function and validate your result/ face as result.

http://www.doc.ic.ac.uk/~dfg/ProbabilisticInference/IDAPILecture15.pdf

https://onlinecourses.science.psu.edu/stat505/node/51

Thank you very much for your valuable suggestions.

Dear Prasad Rao ,

Look the link. May be useful.

http://www.iis.sinica.edu.tw/papers/song/11489-F.pdf

Regards, Shafagat

hi Vandana Agarwal !

          Thank you very much for valuable inputs. I am only the beginner to image processing but i want to learn quickly and looking to work. plz suggest any feature extraction algorithms that can easy to understand and implement.

Hi,

You may start with SIFT and SURF feature extractors. Although they are designed for general purpose of object recognition, yet can be applied on face recognition/retrieval. One example of  face-specific feature extractor is discussed in: Wu, Zhong, et al. "Scalable face image retrieval with identity-based quantization and multireference reranking." Pattern Analysis and Machine Intelligence, IEEE Transactions on 33.10 (2011): 1991-2001.

I would suggest to look into Haar-like Features for facial feature extraction. To start with, this is the most explored and used method in face detection. In OpenCV, you have HaarCascade classifiers for face, eyes, nose and mouth detection.

http://docs.opencv.org/modules/objdetect/doc/cascade_classification.html

you can use these classifiers for detecting/tracking different facial features in your video or still image. Just play around the available sample code in opencv api examples.

Hi, The main approaches for your problem are patch-based method, which they represent the patch and create a model to identify the patch with face. For representing the patch it is possible to use the  Har like feature, SIFT or SURF, and in a advanced way you can follow the deep learning.

if the running time is a important parameter, my suggest is  using cascaded classifier. 

The Cascaded Classifier which is based on Har like is implemented in OpenCV.

For more details see below papers, they are hot paper related your problem. 

A convolutional neural network cascade for face detection.

Fast asymmetric learning for cascade face detection.

The following link, I suppose, would be helpful

http://sprg.massey.ac.nz/ivcnz/Proceedings/IVCNZ_28.pdf