PROJECT ON

STUDY OF FACE BIOMETRY IN

COMPRESSED DOMAIN
Presented by
Dipayan Chatterjee (Roll No.-97/INM/164002)
& Subrata Golui (Roll No.-97/INM/164005)
3Yr. M.Tech in Instrumentation and Control Engineering, 2019
Supervised by
Dr. Saurabh Pal
(Associate Professor)
Applied Physics
University of Calcutta

1
 TOPICS TO BE COVERED
• Aim and Motivation of the Project
• Introduction to General Biometric System
• Compressed Domain Biometry
• Work Flow Diagram of the Project
• Image Database
• Image Preparation
• Compression
• Feature Set Generation
• Comparison
• Result
• Discussion
• Reference

2
 AIM AND MOTIVATION
OF THE PROJECT
 Aim of the project
• Face recognition in compressed domain
 Motivation of the project
• Reduce overall computational time
• Reduce the size of data.

3
 INTRODUCTION
 What is Biometric system?
Stored
Template
Test

Feature Template
Pre-processing Matcher Decision
Extraction Generator Test

Sensor

4
 IMPORTANCE OF BIOMETRIC
RECOGNITION

 Increase Security
 Attendance in office
 Eliminate problems caused by lost IDs or forgotten passwords
 Replace hard-to-remember passwords which may be shared or
observed
 Reduce fraud

5
 BIOMETRIC CHARACTERISTICS

(a) DNA, (b) ear, (c) face, (d) facial thermogram, (e) hand thermogram, (f) hand vein,
(g) fingerprint, (h) gait, (i) hand geometry, (j) iris, (k) palmprint, (l) retina,
(m) signature, and (n) voice.
6
 UNIMODAL BIOMETRICS
 Unimodal Biometric System

Subject

Feature
Feature Sensor Matcher Decision
Extractor

Database

7
 MULTIMODAL BIOMETRICS
Multimodal Biometric System

Sensor Feature
Feature 1 1 extraction Decision

Sensor Feature Feature

Feature 2 2 extraction Fusion Matcher

Sensor Feature
Feature n n extraction Stored
Database

8
 WHY FACE RECOGNITION?
Advantages of Face Recognition as a Biometric are

 Improved Security
 Easy to access and capture
 Fully automated
 Most expressive
 Contains various features

9
 WHY COMPRESSED DOMAIN
BIOMETRY USED?
 It reduces storage requirements
 Reduces detection time
 It also reduces the probability of transmission errors since
fewer bits are transferred
 Removes the redundant information

10
WORK FLOW DIAGRAM OF THE
PROJECT

Database contains uncompressed and 3 level compressed images

11
 DATABASE
 YALE Database
•The database consists of 15 persons.
•Each person has ten different oriented images with
resolution of 240x320.
• We worked with 10 subjects with normal, sad and
happy expressions.

Normal Happy Sad

12
 IMAGE PREPARATION
Database images with resolution of 240 x 320.
Resized it to 256 x 256.
Convert it to gray scale.

13
 IMAGE COMPRESSION USING
SINGULAR VALUE DECOMPOSITION
 Any real m x n matrix A can be decomposed uniquely as:
A = USVT

A = [ u1 u2 u3 ....um ]

The last values of ‘S’ are approximately equal to zero, they can be removed
and A is represented as
A = US1VT

A = [ u1 u2 u3 ....um ]

14
 SVD Contd.
In the above matrix, S values after r terms are approximated to zero.
So multiplication of the terms greater than r will be zero.
If m=n, the matrix A can be represented as

15
 MEMORY UTILIZATION IN SVD
According to definition of SVD, the image matrix can be represented as
A = USVT
After performing SVD, With r approximations, U matrix is of size m x r, V
matrix is of size n x r,

mxn = mxr X rxr X rxn

A U S VT

Before SVD memory After SVD memory space required

space required M A = MU + MS + MV
mxn Or, M A = mr + r + nr

The performance of the SVD image compression can be measure by

Compression Ratio of the compressed image as,
CR =(r (m + n + 1)) / mn
16
 Example of Memory Utilization
Consider A = 100 x 200 size image
Original space require=100x200 pixel
=20000 pixel
For r=32 approximation after SVD
Memory space require= (100x32+32+32x200) pixel
= 9632 pixel
Compression Ratio=9632/20000
= 0.4816
= 0.48(approx)
 So, 48% memory needed than original required.
17
FEATURE EXTRACTION AND FUSION

18
 FEATURE EXTRACTION BASED
ON EIGEN VALUES
For a linear transformation the basic equation
Ax = λx
The scalar λ is an eigenvalue of square matrix A and x is the eigenvector
corresponding to the value of λ.
 For k x k matrix the equation can be written as

 The characteristic equation is of the form

det [A-λI] = 0
 By solving this characteristic equation we can get the eigen values.
19
 FEATURE EXTRACTION BASED
ON EIGEN VALLUE contd.
Eigen values are calculated and sorted in descending order of
magnitude.
From the sorted column matrix we choose 8 maximum eigenvalues.

Image Finding Selection of

Sorting of
Matrix eigen 08 eigen
eigen values
(A) values values

20
 CASCADE DETECTOR BASED
FEATURE EXTRACTION
 The Cascade Detector works on Viola-Jones Algorithm. The key points in
Viola-Jones Algorithm are:
• Haar Features
• Integral Image
• Cascading Classifier
 The Cascade Object Detector function is used to implement the viola-jones
algorithm in matlab.
 After detecting various parts of the face like eyes, nose, mouth, we have
extracted the features using Euclidean distance Method.
 HAAR FEATURES

 Haar features are used to detect the presence of a particular feature or pattern in the
given image based on intensity.
 Each feature results in a single value which is calculated by subtracting the sum of
pixels under white rectangle from the sum of pixels under black rectangle.

Types of Haar features used in Viola Jones Applying on a given image

 INTEGRAL IMAGE
Integral image is an image we get by cumulative addition of intensities on
subsequent pixels in both horizontal and vertical axis.
To calculate average intensity over the area highlighted in blue in matrix A of
size 5 x 5 representing an image normally, we do it as : 9 + 1 + 2 + 6 + 0 + 5 + 3 +
6 + 5 = 37; 37 / 9 = 4.11
This require 9 operations. Doing the same for 100 such
operations would require 900 operations.

Integral image of the image is given. Input Image

Making this image required a total of 56 operations.
Average intensity is calculated as,(76 +5) –(24+20)=37;
37 / 9 = 4.11. This required a total of 4 operations.
To do this for 100 such operations would require
(56 + 100 * 4) = 456 operations. Integral Image
For just hundred operations over a 5 x 5 matrix, using an integral image uses
about 50 % less computations, thus for large images the computation time reduces
drastically.
 CASCADE DETECTOR

 Cascade Detector is pre-trained with many positive sample images of

a particular object and many arbitrary negative images of same size.
 In Viola-Jones face detection we scan the detector many times through
the same image , each time with a new size.

Cascade Detector Showing

2 Stage Cascade Classifier
 CASCADE DETECTOR BASED
FEATURE EXTRACTION
 After processing the image using Cascade Object Based Detector Function
we get several detection of the same part of the face as shown.
 For choosing a single and perfect detection we have designed different
functions for different parts of the face.
 For example if we represent the image by x*y pixel matrix, the actual
mouth detection in a face will have maximum positional value vertically.

Multiple Detections of Mouth Deduced Single Detection of

25
Mouth
 CASCADE DETECTOR BASED
FEATURE EXTRACTION
The distinct parts of a face like eyes, nose, mouth are detected and their dimensions
are found by function step and bbox(or Boundary Box).
This function returns a matrix [x y a b]. x,y,a and b are shown in figure below
where x,y contains positional value of the first pixel of the boundary box while a and b
contains value of width and height respectively.
The features are calculated by using Euclidean distance formula as shown in figure.

26
FINAL FEATURE SET FORMATION

Final feature set

*EV’s- Eigen Value 1, Eigen Value 2,...EigenValue 8

F’s- Feature1, Feature 2, ..Feature 8 27
 COMPARISON
 Comparison is done by calculating root mean square error(RMSE)

 The result depend on the least value of RMSE. 28

 Result
• Compression Ratio = r(m+n+1)/m*n
• mxn is the uncompressed image size and r is the
reduced image dimension in SVD

29
uncompressed 75% compression 50% compression 25% compression
SAMPLE TABLE OF COMPARISON DATA

HT- Happy
Test
NS- Normal
Comparison of 75% Compressed Test Image with 25% Compressed Stored Image Store

Comparison of 50% Compressed Test Image with 25% Compressed Stored Image 30
 Result contd..

Accuracy of detecting or identifying uncompressed Test Images with 3 level

compressed Stored images.

Accuracy of detecting or identifying 75% Compressed Test Images with 31

3 level compressed Stored images.
 Result Contd..

Accuracy of detecting or identifying 50% Compressed Test Images

with 3 level compressed Stored images.

Accuracy of detecting or identifying 25% Compressed Test Images

32
with 3 level compressed Stored images.
 DISCUSSION
 Accuracy of normal face image detection is not hampered
by compressing the images. It is always 100%.
But in case of happy and sad posed faces, image detection
accuracy is 80% up to 75% image compression. It is also to
be noted that the accuracy is reduced to 70% in some cases
of 50% compression and all cases of 25% compression.
The reduction in accuracy happens due to changes in
facial expression which causes facial dimension change So,
in turn some features also change.

33
 Reference
• [1] A. Lanitis, C. J. Taylor, and T. F. Cootes, “An automatic face identification system using
flexible appearance models,” Image and Vision Computing, vol.13, no.5, pp.393-401, 1995.
• [2] I. Craw, D. Tock, and A. Bennett, “Finding face features,” Proc.of 2nd European Conf.
Computer Vision. pp. 92-96, 1992.
• [3] T. K. Leung, M. C. Burl, and P. Perona, “Finding faces in cluttered scenes using random
labeled graph matching,” Proc. 5th IEEE int’l Conf. Computer Vision, pp. 637-644, 1995.
• [4] B. Moghaddam and A. Pentland, “Probabilistic visual learning for object recognition,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no.7. pp. 696-710, July, 1997.
• [5] M.A. Turk and A.P. Pentland, “Face Recognition Using Eigenfaces”, IEEE Conf. on
Computer Vision and Pattern Recognition, pp. 586-591, 1991.
• [6] The Yale Face Database--http://vision.ucsd.edu/content/yale-face-database
• [7] P. Viola and M. Jones, “Robust real-time face detection,” Computer Vision, 2001. ICCV
2001. Proceedings. Eighth IEEE International Conference on, vol. 2, pp.747, 2001.

34