Table 1 Different feature extraction algorithms on FERET data sets

1

Gabor PCA [19]

\( GP = argmax(Real\left( {I*\psi } \right)^{2} + Img\left( {I*\psi } \right)^{2} )*\lambda \)

Extracted real and imaginary features of Gabor (G) were optimized by PCA (P)

Extracts edges from the edges by different frequency scaling components concerning PCA provides a recognition rate of 89.4%

2

Gabor MEAN PCA [19]

\( GP = argmax(\sum (Real\left( {I*\psi } \right)^{2} + Img\left( {I*\psi } \right)^{2} ) ) *\lambda \)

Here mean of Gabor filter were subjected to PCA optimization

All the extracted edges applied by mean and provide a recognition rate of 92.7%

3

HOG [8]

\( H = \frac{v}{{\sqrt {v_{2}^{2} + e^{2} } }} \)

This theory will perform local feature extraction and normalization.

All the block-wise extracted histogram features will result in an under-recognition rate of 83.9%

4

SIFT [12]

Scale

\( D\left( {x,y,\sigma } \right) = L\left( {x,y,K_{i} \sigma } \right) - L\left( {x,y,K_{j} \sigma } \right) \)

Position:

\( D\left( x \right) = D + \frac{{\partial D^{T} }}{\partial x}x + \frac{1}{2}x^{T} \frac{{\partial^{2} D}}{{\partial x^{2} }}x \)

Orientation:

\( m\left( {x,y} \right) = \sqrt {\left( {L\left( {x + 1} \right) - L\left( {x - 1} \right)} \right)^{2} + \left( {L\left( {y + 1} \right) - L\left( {y - 1} \right)} \right)^{2} } \)

\( \theta \left( {x,y} \right) = atan2\left( {L\left( {y + 1} \right) - L\left( {y - 1} \right),L\left( {x + 1} \right) - L\left( {x - 1} \right)} \right) \)

One descriptor base four features will be extracted from the SIFT algorithm, and point descriptor of the best form will help exact identification of objection. Even the scale of the object varied yet will recognize it.

As per remarks, it compares the object with point key descriptors and results with an accuracy of 92.7% for recognition

5

MSIFT-Proposing

Scale

\( D\left( {x,y,\sigma } \right) = \left( {L\left( {x,y,K_{i} \sigma } \right) - L\left( {x,y,K_{j} \sigma } \right)} \right)*\varPsi \)

Position:

\( D\left( x \right) = \left( {D + \frac{{\partial D^{T} }}{\partial x}x + \frac{1}{2}x^{T} \frac{{\partial^{2} D}}{{\partial x^{2} }}x} \right)*\varPsi \)

Orientation:

\( m\left( {x,y} \right) = \sqrt {\left( {L\left( {x + 1} \right) - L\left( {x - 1} \right)} \right)^{2} + \left( {L\left( {y + 1} \right) - L\left( {y - 1} \right)} \right)^{2} } \)

\( \theta \left( {x,y} \right) = atan2\left( {L\left( {y + 1} \right) - L\left( {y - 1} \right),L\left( {x + 1} \right) - L\left( {x - 1} \right)} \right) \)

The image is split into multiple frequencies components and extracted sift features from them will result in effective detection even in images of different sizes.

The modified SIFT results in most format and results in the best recognition rate of 98.3% in this proposing approach