hi sender thanks for suggestion . however this  task is specific. please  I need your help. 

Hi. ..

May be this will help you....

http://biometrics.idealtest.org/

As a simple first pass, I would just concatenate all of the projections together into a single feature vector and pass it through whatever type of classifier you want to use.

Yes Mathew, i thought of that, but there is still one challenge, remember the projections from the five methods will not be of equal size considering that for spPCA and mPCA we divide the images into different partition for training as compared to other global methods like PCA and LDA. Any advice on this?

I'm not familiar with these two methods except for a cursory reading just now.  It seems like they divide the face images into multiple parts and then run PCA separately in each region.  Again, as a simple first pass, I would just try to concatenate all the vectors together, including the separate projections from each region.  The Histogram of Oriented Gradients (HOG) feature extraction did something similar and even though it's an old method it still has competitive performance on pedestrian detection.

I suggest you this:

create a template and assign a number of bits for each feature vector of each method. In this way you can read and write as many feature vector of many methods as you want and you always will be able to read and know which feature vector you are reading.

Let me know if you understand.

Best regards

I suggest to use Particle Swarm Optimization (PSO) to select the best subsets of face and  Iris features.

Best regards.

Hi. I would concatenate the feature vectors into one long feature vector, and then perform classification using the combined feature vector. You first need to normalize the individual modalities, to even out the distribution.

This paper here describe how they solve face and iris feature level fusion

http://www.jcomputers.us/vol6/jcp0605-13.pdf

Thanks chiung i will look into this and get back to you

Biometric fusion can be done on 3 levels:

1- Feature Level Fusion: concatenate both Iris and face feature vectors and carry on training/testing as normal, you may add some weights to Iris as it is more accurate than the face. remember to normalize all the attributes after fusion using Min-max, z-score, etc. you may use PCA , LDA for dimnensionality reduction 

2- Matching level: each method work alone, the matching results are taken using either, sum rule (normal summation of the matching results), minimum score, maximum score or Matcher Weighting (give the highest weight to the method of the least equal error rate)

3- decision level: each method works alone, on decision you may use: majority voting, complete agreement, or Weighted Sum Rule.

https://www.researchgate.net/publication/297245624_Particle_Swarm_Optimization_Algorithm_and_its_Codes_in_MATLAB

Research Particle Swarm Optimization: Algorithm and its Codes in MATLAB

https://www.researchgate.net/publication/296636431_Codes_in_MATLAB_for_Particle_Swarm_Optimization

Code Codes in MATLAB for Particle Swarm Optimization