Techniques Used in Face Recognition

• Dimensionality Reduction Approaches:

Dimensional difficulty is the most significant problem of Face Recognition. In this Approaches the methods are beneficial to reduce the dimensions of the considered space. This issue is reduced by dimensionality reduction techniques. Among those Techniques some are linear and some are non-linear. Some of the Dimensionality reduction techniques are as follows:

1 Principal Component Analysis (PCA):

One of the most important technique that is used for image recognition and compression is the Principal Component Analysis (PCA).PCA was applied for representation of face images (Sirovich et al 1987). PCA was extended to recognize faces from Eigen faces in this the size of training image and test image must be same. (TURk et al 1991)PCA uses an orthogonal technique for converting set of correlated variables into set of uncorrelated variables. That uncorrelated variables are called Principal Components. To express 1D vector of pixels from 2D image into the compact principal components of the feature space is the main purpose of PCA. This process is called Eigen space projection. The calculation of Eigen space can be done by finding the Eigen vectors of facial image using covariance matrix. Normally to choose a threshold is very difficult. The main aim in PCA is to reduce high dimensionality of the original dataset. A small set of Variables which are Uncorrelated are much easier to understand and can be used for further analysis then the original data sets where the variables are correlated. PCA is used for pattern identification in the original data set. It expresses data in a way that similarities and differences are highlighted. Once these patterns are found the data can be compressed.(Smith at al 2002).

There are number of steps in PCA:

1) Get the original data

2) Calculate the mean from the original data

3) Subtract the mean from the original data.

4) From mean subtracted data calculate covariance matrix.

5) From covariance matrix calculate Eigen-Values and Eigen-Vectors.

6) Choose the Principal Components from the featured space.

7) Derive the new data set.

PCA is also called Karhunen Loeve Transform (KLT) method. It is one of the widely used method for pattern recognition, computer vision etc. (Yang et al 20014).

Advantages of PCA

Simplest approach used for data used for face recognition and data compression.

Operating rate is fast.

Dis Advantages of PCA

Full Frontal display of faces is required

Insensitive to position of face and lighting condition.

Every face in this is considered as a different image

2 Independent Component Analysis (ICA):

Independent Component Analysis also comes under dimensionality reduction approaches. It is also called blind source separation technique. ICA finds the projection directions of independent components such that data projected onto these directions have maximum statistical independence. ICA was primary used for signal processing. Here we can give an example let we are in a room where there are three persons and they are speaking simultaneously (numbers are supposition otherwise it could be any number). We have also three microphones, which are located at different locations, they gave three recorded time signals, which can be denoted as:

Y1 (t), Y2 (t) and Y3 (t).

Where Y1, Y2 and Y3 are the amplitudes of signal and t is index of time.

The three speaker which emitted the speech signal are recorded as the weighted sum which can be denoted as:

C1 (t), C2 (t) and C3 (t).

This could be expressed as linear equation as:

Y1 (t) = A11 C1 (t) + A12 C2 (t) + A13 C3 (t).

Y2 (t) = A21 C1 (t) + A22 C2 (t) + A23 C3 (t).

Y3 (t) = A31 C1 (t) + A32 C2 (t) + A33 C3 (t).

Where Aij with i=1…3 and j=1..3 are the parameters that depend on the microphones distance from the speaker.

To estimate the original signal using the recorded signal is called a cocktail problem.(Yang et al 2002)

The ICA technique allows us to separate the two or more than two original source signal from there mixture. This technique has a lot of applications in image processing, pattern recognition, image separation etc. ICA can be seen as PCA’s generalization. The BSS model in ICA is as Follows (Darper 2003).

3 Linear Discriminant Analysis (LDA):

LDA is also comes under dimensionality reduction approaches. LDA has showed great performance in the face recognition.