Sift Findas

8/3/2019 Sift Findas

1/32

DEPARTMENT OF INFORMATION SCIENCE & ENGINEERING

FACE RECOGNITIONUSING SIFT FEATURES

Presented by,

Nouman Sadiq(1PI08IS055)

Deep Agarwal (1PI08IS034)

Department of ISE

PESIT,

Bangalore

Internal guide,

Prof. Shylaja S SH.O.D

Department of ISE

PESIT,

Bangalore

1


2/32

AGENDA

2

INTRODUCTION1

PROJECT BRIEF2

IMPLEMENTATION3

4 CONCLUSSION

5 FURTHER ENHANCEMENT

6 BIBLIOGRAPHY

Aug 2012 Dec 2012 Department of Information Science Eng.


3/32

PROJECTOVERVIEW

Scope of this project?

Distinctive invariant features from a

face is obtained. It is applied to the training set of the

faces, thus transforming it.

The face in the test set is matched.3

Goal of this project?

To extract Scale Invariant Featuresfrom a face that can be further used

to perform reliable face recognition.



4/32

Face recognition can be categorized into three steps

Face detection

Feature extraction

Face recognition

Face recognition is always prone to problems like

Change of posture

Illumination changes

Change of environment

4

Face recognition??



5/32

Stands for Scale Invariant Feature

Tranformation

David G. Lowe Introduced this algorithm.

Here image features having properties which

makes them suitable for matching differingimages of the same face are extracted.

5

Introduction

- S.I.F.T



6/32

SIFT ALGORITHM OVERVIEWFollowing are the major stages of computation used to

generate the set of image features:

1. Scale-space extrema detection

2. Keypoint localization

3. Orientation Assignment

4. Keypoint Descriptor construction

6



7/32

It involves the following steps

1.1 : The face is expressed as octaves ofdifferent size. I(x,y)

1.2 : Each images of an octave issmoothed with different scales ofthe Gaussian function. L(x,y)

1.3 : Compute difference of Gaussian(doG) images from adjacent scalesfor entire octave D(x,y).

1.4 :From difference-of-Gaussian localextrema detection we obtainapproximate values for keypoints (orinteresting points)

7

The input image

The input image

with interesting

points


STEP1:SCALE-SPACEEXTREMADETECTION


8/32


1.1 : The face is expressed as octaves of different size

8


Four Different Octaves


9/32

1.2 : All the octave images are filtering with Gaussianfunction thus obtaining different scales.

9

Where

The scale space of image L(x, y, ), that is produced from

the convolution of a variable-scale Gaussian, G(x, y, )withan input image, I(x, y):




10/32

1.2 : All the octave images are filtering with Gaussianfunction thus obtaining different scales.

10The second Octave of L images


Five Different scales of

Gaussian blur()



11/32

1.3 : The difference of the different scale images isfound.

11The second Octave of D images




12/32

1.3 : The difference of the different scale images isfound.

12




13/32

13




14/32

1.4 :From difference-of-Gaussian local extrema detectionwe obtain approximate values for keypoints

14

Maxima and minima of thedifference-of-Gaussianimages are detected bycomparing apixel (marked with X) to its26 neighbours in 3x3

regions at the current andadjacent scales (markedwith circles)




15/32

1.4 :From difference-of-Gaussian local extrema detection weobtain approximate values for keypoints

15




16/32

STEP 2:KEYPOINTLOCALIZATION

It involves the following steps

2.1 : The interesting points of very Low

contrast are removed

2.2 : Some more points are removedwhich threshold on ratio of principalcurvatures.

16

The image with the

interesting points

The image with the

keypoints

Here the keypoints are selected basedon measures of their stability.



17/32


17

2.1 : The interesting points of very Lowcontrast are removed

The image with the

interesting points

The image with the

keypoints



18/32


18

2.2 : Some more points are removed which thresholdon ratio of principal curvatures.

Difference-of-Gaussian function will be strong

along edgesSome locations along edges will have a large

principal curvature across the edge but a smallprincipal of curvature perpendicular to the edge

Therefore we need to compute the principalcurvatures at the location and compare the two.

Then eliminate some of the candidates belowthreshold



19/32


19

The image with

interesting points

The image with the

keypoints



20/32


20

436 interesting

points

The points below

contrast threshold is

eliminated.410

points eliminated

26 interesting

points are left

Further points

whose curvature

ratio is above the

threshold is also

removed. 13 more

points are removed.

The final face with

the 13 keypoints.



21/32


22/32

STEP 3:ORIENTATIONASSIGNMENT

For image sample, L(x, y), the gradient magnitude,

m(x, y), and orientation, teta(x, y), is computed usingpixel differences:

22

Points in region around keypoint are selected andmagnitude and orientations of gradient arecalculated.

22))1,()1,(()),1(),1((),( yxLyxLyxLyxLyxm

))),1(),1(/())1,()1,(((tan),(1

yxLyxLyxLyxLyx

Aug 2012

Dec 2012 Department of Information Science Eng.


23/32

TRAININGFACESET

23

Aug 2012



24/32

TESTFACESET

24

Aug 2012



25/32

EXTRACTING KEYPOINTS

25

Aug 2012


Video 1


26/32

EXTRACTING KEYPOINTS

26

Aug 2012


Video 2


27/32


28/32

28

Aug 2012


Two images of the faces of a different persons.

0 keypoints are matched

MATCHING


29/32

29

CONCLUSION

In this project Scale Invariant Feature Transform (SIFT) is

implemented for feature extraction for the purpose of face

recognition.

The Comparisons of this approach among other holistic

approaches and feature based approaches are yet to be

done .

Aug 2012



30/32

30

Further Enhancement

There are many directions for further research in deriving

invariant and distinctive face Features, further distinctiveness

could be derived from including illumination-invariant color

descriptors .

Another direction for future research will be to individually

learn features that are suited to recognizing particular

categories of facial images.

Aug 2012



31/32


32/32

32Aug 2012 Dec 2012 Department of Information Science Eng.

Sift Findas

Documents