You are on page 1of 24

CL311

Technical Wri ng and Presenta on


Department of Chemical Engineering

A Project Report on
FACE RECOGNITION SYSTEM

Submi ed by :
Rishabh Lal (200107066)
Adarsh Kumar (200107004)
Ashutosh Kumar (200107011)
ACKNOWLEDGMENT

We are grateful to the Almighty for gran ng us the for tude and
persistence to finish the main project. The ini a ve itself serves as a thank you
to everyone who has donated their hear elt co-opera on in making this
project a grand success.

We are greatly indebted to project guide Mr. Mihir Purkait,Professor and Mr.
Anweshan, Research Scholar, Chemical Engineering, for providing valuable
guidance at every stage of the project work.We are profoundly grateful towards
the unmatched services rendered by him.

Finally, but certainly not least, we would like to express our sincere gra tude
and apprecia on to our beloved parents for their spiritual support and genuine
assistance in carrying out the primary project.

IIT GUWAHATI
ABSTRACT

One of the simplest ways to tell someone's iden ty from another is by


looking at their face.A personal iden fica on method called face recogni on
analyses a person's physical features to determine their iden ty. The process of
recognising a face mainly consists of two phases: face detec on, which occurs
quickly in people, excluding situa ons when the item is nearby. The second
step is introduc on, which iden fies a face as belonging to a certain person.
Then, the stage is duplicated and used as a model for facial image recogni on
(also known as "face recogni on"), a biometrics technique that has been
extensively researched and created by professionals.

There are two different types of approaches that have been created for
facial recogni on pa erns, the Eigenface method and Fisherface method.
Image recogni on of face : The Eigenface approach is based on the Principal
Component Analysis (PCA) of facial data to reduce the face-dimensional space.
Facial space (face recogni on u lizing Eigen faces) was created for the primary
usage of PCA by iden fying the facial image's eigenvector that corresponds to
its biggest eigenvalue. the region of image processing is being used in this
project's face detec on and face recogni on systems.

IIT GUWAHATI
INDEX

CONTENTS Page No.

1. INTRODUCTION……………………………………………………………………………..4
1.1 FACE RECOGNITION......................................................................4
1.2 FACE DETECTION..........................................................................6
2. FACE DETECTION METHODS.............................................................8
2.1 FEATURE BASED APPROACH........................................................9
2.1.1 SNAKES...............................................................................9
2.1.2 DEFORMABLE TEMPLATES.................................................10
2.2 LOW-LEVEL ANALYSIS..................................................................10
2.3 MOTION BASE.............................................................................12
2.3.1 GRAY SCLAE BASE..............................................................12
2.4 FEATURE ANALYSIS......................................................................13
2.4.1 VIOLA JONES METHOD......................................................13
3. FACE RECOGNITION……………………………………………………………………..15
3.1 FACE RECOGNITION USING GEOMETRICAL FEATURES……………..15
3.1.1 FACE RECOGNITION USING TEMPLATE MATCHING…………16
3.2 PROBLEM SCOPE AND SYSTEM SPECIFICATIONS……………………..16
3.3 DIFFICULTIES IN FACE RECOGNITION……………………………………….17
3.4 IMPROVING FACE DETECTION USING RECONSTRUCTION………..18
4. PROBLEM SOLVED USING FACE RECOGNITION SYSTEM……………….20
5. CONCLUSION……………………………………………………………………………….22
6. REFERENCE………………………………………………………………………………….23

IIT GUWAHATI
4

1. INTRODUCTION

Finding out whether a previously observed object is a known or unknown


face is known as face recogni on. Frequently, the issue of face detec on is
mistaken with the issue of face recogni on. Face Recogni on, on the other
hand, uses a database of faces to authen cate the input face and determine
whether the "face" is a known or unknown person.

1.1 FACE RECOGNITION:


DIFFERENT APPROACHES OF FACE RECOGNITION:
There are two main methods for tackling the issue of face recogni on:
geometric (based on features) and photometric (view based). Three of the
many algorithms created as a result of the growing interest in face recogni on
among researchers have been extensively researched in the face recogni on
literature.

Recogni on algorithms can be divided into two main approaches:

1. Geometric: is based on the spa al arrangement of facial features, or


the geometric rela onship between facial landmarks. This implies that faces are
first classified according to different geometrical angles and distances between
features before the key geometrical characteris cs of the face, such as the eyes,
nose, and mouth, are located in Figure 1.

IIT GUWAHATI
5

Figure 1: Geometric Facial Recogni on

2. Photometric stereo: Used to reconstruct an object's shape from a


collec on of photos taken in various ligh ng situa ons. An array of surface
normals comprise a gradient map, which determines the geometry of the
recovered item (Figure 2).

Figure 2: Photometric Stereo Image

IIT GUWAHATI
6
Popular recogni on algorithms include:
1. Principal Component Analysis using Eigenfaces, (PCA)
2. Linear discriminant analysis,
3. Elas c Bunch Graph Matching using the Fisherface algorithm,

1.2 FACE DETECTION:


Separa ng image windows into two classes, one of which contains faces
(while tarnishing the background), is required for face detec on (clu er). It is
challenging because, despite the similari es across faces, they might differ
greatly in terms of age, skin tone, and facial expression. Different ligh ng
situa ons, picture quality, and geometries, as well as the poten al for par al
occlusion and disguise, make the issue even more difficult to solve.

Therefore, the perfect face detector would be able to iden fy any face
against any background, in any ligh ng situa on. There are two steps to the
face detec on task. In the first stage, a classifica on task is used to determine
whether there are any faces in an input image. This task accepts any image as
input and outputs a binary value of yes or no. The face localiza on task, which
is the second phase, outputs the loca on of any face or faces inside an image
as a bounding box containing (x, y, width, height).

The face detec on system can be divided into the following steps:-
1. Pre-Processing: Before being sent into the network, the photos are
processed to lessen the variability in the faces. We collected all instances that
are favourable by cropping the photographs of faces. Frontal faces on page 5
should only be seen from the front. The clipped photos are then all
light-adjusted using industry-standard methods.
2. Classifica on: Neural networks are implemented to classify the images
as faces or nonfaces by training on these examples. We use both our
implementa on of the neural network and the Matlab neural network toolbox

IIT GUWAHATI
7

for this task. Different network configura ons are experimented with to
op mise the results.
3. Localiza on: When looking for faces in an image, the trained neural
network locates them within a bounding box. Several facial features that have
been studied include: Posi on Scale Orienta on Illumina on.

Figure 3: Face Detec on Algorithm

IIT GUWAHATI
8

2. FACE DETECTION METHODS

Face detec on is a computer technology to locate and quan fy human


faces in random (digital) images. Any other objects in the digital image,
including bodies, trees, and buildings, are ignored in favour of the facial
features.
Face localiza on, or face detec on, can be thought of as a more "generic"
situa on. The goal of face localiza on is to determine the loca ons and
dimensions of a known number of faces (usually one). The two main methods
for detec ng facial parts in a given image are feature-based and image-based
methods. The feature-based technique makes an a empt to extract image
features and compare them to the features of a face that are known. In
contrast, the image-based technique seeks the best alignment between training
and tes ng images.

Figure 4: Face Detec on Methods

IIT GUWAHATI
9

2.1 FEATURE BASED APPROACH:


Ac ve Shape Models Focus is placed on complicated, non-rigid elements
like actual physical characteris cs and higher level appearances of
characteris cs in ac ve shape models translates to mean that Ac ve Shape
Models (ASMs) are intended to automa cally locate landmark points that
determine the shape of any sta s cally modelled object in an image. Consider
the appearance of the eyes, lips, nose, mouth, and eyebrows. Building a
sta s cal model is a component of an ASM's training stage.
2.1.1 Snakes:
Snakes are used to iden fy head boundaries [8,9,10,11,12]. In order to
achieve the task, a snake is first ini alized at the proximity around a head
boundary. It then locks onto nearby edges and subsequently assume the shape
of the head.
Snakes have some drawbacks, including the fact that its contour
frequently gets caught on erroneous image features and that they are
unsuitable for extrac ng non-convex features.(Figure 5).

Figure 5: Snake Face Detec on Method

IIT GUWAHATI
10
2.1.2 Deformable Templates:
In order to improve the performance of snakes and account for the a
priori of facial features, Yuille et al. proposed deformable templates. It is
challenging to arrange the local evidence of facial edges into a logical global
en ty using generic contours, making it difficult to locate a facial feature
boundary. The edge recogni on technique is complicated by the low brightness
contrast surrounding some of these objects. By using global informa on from
the eye to increase the accuracy of the extrac on process, Yuille et al.
expanded on the idea of snakes.
Approaches using deformable templates are created to address this issue. Local
valley, edge, peak, and brightness are the basis for deforma on. The extrac on
of salient features (eyes, nose, mouth, and brows), in addi on to the facial
border, is a significant challenge in face recogni on. E = Ev + Ee + Ep + Ei +
Einternal, where Ev, Ee, Ep, Ei, Einternal are internal energy and exterior energy
resul ng from the valley, edges, peak, and image brightness.

Figure 6: Deformable Templates methods

2.2 LOW LEVEL ANALYSIS:


Based on basic visual characteris cs, such as colour, intensity, edges,
mo on, etc. Base Skin Color A crucial aspect of human faces is colour. There are
many benefits to monitoring a face using skin tone as a characteris c.
Processing of colours occurs considerably more quickly than processing of other

IIT GUWAHATI
11

facial aspects. Colour is orienta on invariant under certain illumina on


condi ons. Due to the fact that only a transla on model is required, this trait
greatly simplifies mo on es ma on. The colour representa on of a face
received by a camera is impacted by several elements (environmental light,
object movement, etc.), making it difficult to track human faces using colour as
a feature.

Figure 7: Low Level Analysis

Majorly three different face detec on algorithms are available based on


RGB, YCbCr, and HIS color space models.In the implementa on of the
algorithms there are three main steps viz.

1. Classify the skin region in the color space,


2. Apply threshold to mask the skin region and
3. Draw bounding box to extract the face image.

This method also has the restric on that the skin region in the image
must only be on the face. To dis nguish skin patches from the backdrop in the
HSV colour space, Kjeldson and Kender created a colour predicate.

IIT GUWAHATI
12
The classifica on of skin colour in HSI colour space is iden cal to that in
YCbCr colour space, however here the relevant values are hue (H) and
satura on (S). A pixel is categorised as having skin tone if the values [H,S] lie
inside the threshold, similar to above the threshold be chosen as [H1, S1] and
[H2, S2], and this distribu on provides the localised facial image. This algorithm
is subject to the same constraint as the previous two.

2.3 MOTION BASE:


The loca on of moving objects can be determined using mo on
informa on when a video sequence is provided. Simple thresholding of the
cumula ve frames differences can be used to extract moving silhoue es like
face and body components. Frame differences can be used to find facial
characteris cs in addi on to face regions.

2.3.1 Gray Scale Base:


A face's grey areas might also be treated as significant features. In
general, facial features like lips, brows, and pupils appear darker than the rest
of the face. Within segmented face regions, a number of new feature extrac on
techniques look for local grey minima.
To enhance the quality of local dark patches and facilitate detec on, these
algorithms first apply contrast-stretching and gray-scale morphological
processes to the input images. Dark-patches are extracted using low-level
gray-scale thresholding. based approach with three ers. Yang and Huang
introduced a novel strategy, focusing on how faces behave in pyramidal
(mosaic) images. In this system, hierarchical There are three layers to a face
loca on. Based on mosaic photos of varying resolu ons, higher two level. A
method for edge detec on is put forth at the bo om level.(Figure 8).

IIT GUWAHATI
13

Figure 8: Gray Scale Images

2.4 FEATURE ANALYSIS:


These algorithms seek to iden fy structural characteris cs even when the
purpose, perspec ve, or use the various ligh ng se ngs to find faces. These
techniques are primarily intended for face localisa on.
2.4.1 Viola Jones Method:
A method for object detec on that lowers compu ng me while
obtaining excellent detec on accuracy was presented by Paul Viola and Michael
Jones. A quick and reliable face iden fica on method with 95% accuracy at
about 17 frames per second was proposed by Paul Viola and Michael Jones. It is
15 mes faster than any methodology available at the me of release.
The method depends on the usage of straigh orward Haar-like features that
are swi ly assessed with the aid of a new image representa on. Based on the
idea of a "Integral Image," it generates a huge collec on of features, reduces
the overcomplete set using the boos ng algorithm AdaBoost, and introduces a
degenera ve tree of the boosted classifiers to allow for rapid and reliable
interferences.

IIT GUWAHATI
14

Figure 9: Viola Jones Method

IIT GUWAHATI
15

3. FACE RECOGNITION

Over the previous couple of many years many strategies were proposed
for face reputa on. Many of the strategies proposed throughout the early
ranges of laptop imagina ve and prescient can't be taken into considera on
successful, however nearly all the latest methods to the face reputa on trouble
were creditable. According to the studies through Brunelli and Poggio (1993) all
methods to human face reputa on may be divided into strategies: (1)
Geometrical func ons and (2) Template matching

3.1 FACE RECOGNITION USING GEOMETRICAL FEATURES


Using the image of the face we wish to recognise, this technique
calculates a collec on of geometric a ributes like nose width and length,
mouth posi on, chin shape, etc.
The characteris cs of known people are then compared to this collec on
of characteris cs. The closest match can be determined using a suitable metric,
such as Euclidean distance (finding the closest vector). Although Craw et al.
(1987) conducted rela vely recent research in this field, geometric features
were mostly used in the early face recogni on research (Kanade, 1973).
The benefit of employing geometrical features as the founda on for face
recogni on is that recogni on is achievable even with noisy photos (images
with mul ple disordered pixel intensi es) and at very low resolu ons. The
face's general geometrical configura on can be derived for face iden fica on
even while the face cannot be seen in detail. The primary drawback of the
method is how difficult it is to automate the extrac on of facial geometrical
elements. Scaling and rota on of a face in the image plane have a significant
impact on automated geometrical feature extrac on-based recogni on
(Brunelli and Poggio, 1993). When we look at Kanade's (1973) findings, where
he reported an iden fica on rate of between 45-75% with a database of only
20 persons, this becomes clear.

IIT GUWAHATI
16

3.1.1 Face recogni on using template matching


This is similar to the template matching technique used in face detec on,
except that in this case, our goal is to iden fy a face rather than categorise a
picture as a "face" or "non-face."
Regions of the en re face, including the eyes, nose, and mouth, that
could be employed in a template matching technique. The template matching
strategy's founda on is the extrac on of en re facial areas (a pixel matrix) and
comparison of those with the saved photos of well-known people. The closest
match can once more be determined using Euclidean distance. Baron employed
the straigh orward method of comparing grey-scale intensity values for face
iden fica on (1981). There are, however, a lot more advanced approaches to
template matching in face recogni on.
These entail laborious extrac on of the extracted grey-level intensity
values and transforma on. For instance, Turk and Pentland (1991a)
pre-processed the grey levels using Principal Component Analysis, also referred
to as the eigenfaces approach, while Wisko et al. (1997) pre-processed the
retrieved regions using Elas c Graphs encoded with Gabor filters. Brunelli and
Poggio (1993) compared geometrical features and template matching for face
recogni on and came to the conclusion that while a feature-based approach
might provide faster recogni on and less memory usage, template-based
strategies provide higher recogni on accuracy.

3.2 PROBLEM SCOPE AND SYSTEM SPECIFICATION


A er studying the research on face detec on and recogni on and
considering poten al real-world applica ons for such systems, the problem
scope for this project was established. The ensuing system (systems)
requirements were found
1. A method for spo ng frontal view faces in s ll pictures
2. A method for iden fying a frontal view face.
3. The face detec on and iden fica on system will only be shown frontal view,
expressionless faces.

IIT GUWAHATI
17

4. A high level of ligh ng invariency must be demonstrated by all implemented


systems.
5. Near real- me performance is a need for all systems.
6. Support must be provided for both fully automated and manual face
detec on.
7. Frontal view face recogni on will only require one exis ng image.
8. Automated face detec on and recogni on systems ought to be merged into
one completely automated system. The segmented image generated by the
face detec on sub-system must exhibit some degree of invariency to scaling
and rota on faults in the face recogni on sub-system.
9. A posi on invariant face recogni on system ought to be added to the frontal
view face recogni on system.
Unfortunately, even if we give our problem domain restric ng
requirements, it might not be prac cal to implement a system in the real world
precisely according to these condi ons.

3.3 DIFFICULTIES IN FACE RECOGNITION

1. Recognize comparable faces (inter-class similarity).


2. Facial equipment
3. Effects of ageing
4. Animated faces

It's possible for people to look remarkably similar to one other.


In addi on to other biometric technologies like fingerprint, signature, re na,
and so forth, the face recogni on and detec on system is a pa ern recogni on
method for personal iden fica on reasons. Before being sent into the network,
the photos are processed due to the variety in the faces. Cropping frontal face
photographs to only include the front view yields all of the favourable
examples, which are the face images. The clipped photos are all then
light-adjusted using industry-standard methods.

IIT GUWAHATI
18
faces with intra-subject differences in ligh ng, expression, accessories, colour,
occlusions, and brightness.

3.4 IMPROVING FACE DETECTION USING RECONSTRUCTION


Since reconstruc on would require resizing the face detec on window
area and significant matrix mul plica on, both of which are computa onally
expensive, it cannot be u lised to detect faces in photos in near real- me.
Reconstruc on, on the other hand, can be used to confirm whether genuine
face posi ons suggested by the deformable template method are there. The
face detec on window is probably devoid of a face if the reconstructed image
significantly differs from it. The face detec on method can be adjusted to
generate several high 'faceness' loca ons that can be verified via reconstruc on
rather than just one poten al face posi on. This is especially helpful since
occasionally the op mal frontal view face pixel area may not be present in the
best 'faceness' loca on determined by the deformable template technique.

The face iden fica on system checks prospec ve face loca ons it has
discovered (the best face loca ons it could find during its search) to see if a face
is there. This method of iden fying a face will work effec vely if the threshold
level (the largest difference between the reconstruc on and the original for the
original to be a face) is set correctly. The search space of probable face
placements can be condensed to a small number of spots by the quick
deformable template technique.
Then, a reconstruc on is used to verify these. By instruc ng the face
detec on system to output any loca on with a "faceness" value that is, for
example, at least 0.9 mes the best heuris c value that has been found thus
far, rather than just the best face loca ons it has discovered thus far, it is
possible to alter the number of loca ons the system finds. The next step is to
use reconstruc on to assess a large number of addi onal probable face sites.
Making judgments about these and other speed versus accuracy trade-offs
requires taking into account the pla orm that the system will be used on.

IIT GUWAHATI
19

The output of the face detec on system can be examined for correla on
with the typical face rather than reconstruc on to ensure accuracy. There is
probably a face in the divided areas with a high correla on. Once more, a
threshold value will need to be set in order to dis nguish between faces and
non-faces. Like reconstruc on, resizing the segmented area and figuring out its
associa on with the average face is far too expensive to be u lised by itself for
face detec on, but it is acceptable for confirming the results of the face
detec on system.

IIT GUWAHATI
20

4. Problem Solved Using Face Recogni on System

Today, security is everyone's first priority. Security is necessary


everywhere, including in residen al, business, governmental, and defence
se ngs. The security-provided or accessing system has evolved over me to
include password access, mechanical system-based access, and access based on
biometric features. Fingerprint, iris, and face recogni on and detec on are
frequently u lised in biometric-based systems.
Different discrete access codes have been used in security systems over the
years, but more recently, facial recogni on concepts or ideas have been used to
a variety of challenges like criminal iden fica on and crowd surveillance.
One of the instruments that informa on technology (IT) is a emp ng to fully
u lise is biometric security. Security is one of the biggest problems we currently
face in our daily lives; there are many different security systems, the majority of
which are electronic or manual. However, biometric security check systems,
which are specific to each person, are the most important. Biometric tools
include iris, fingerprints, palmcheck, and facial recogni on, for instance.
An effec ve facial recogni on system can be very useful for forensic sciences,
law enforcement iden fica on, surveillance, banking authen ca on, and
security systems. Face recogni on technology can be used to restrict or prevent
the drawbacks and flaws of human-based security systems.

Among biometric systems, face recogni on is a common one. In essence,


these systems recognize pa erns by retrieving biometric informa on about a
person, extrac ng a feature set from that data (in this case, the face), and
comparing that feature to a template set stored in the database.

A biometric system can func on in iden fica on mode or verifica on


mode, depending on the applica on context. In recent years, face detec on has
become increasingly important. This technology is used by law enforcement

IIT GUWAHATI
21

agencies, security agencies, systems for individual iden fica on, and
applica ons for widespread surveillance. Due to various aspects, including face
size, image size, type, and other circumstances involved, human face detec on
has historically been a difficult undertaking. Faces may be depicted in
photographs in a variety of posi ons and perspec ves, and there are wide
varia ons across photographs and images in terms of quality and contrast.

IIT GUWAHATI
22

Conclusion

In the past 20 years, face recogni on technology has advanced significantly.


Today, automated iden fica on informa on verifica on is possible for safe
transac ons, security and surveillance opera ons, and building access control,
among other applica ons. Because these applica ons typically operate in
restricted contexts, recogni on algorithms can benefit from those limita ons
and achieve high recogni on accuracy. Next-genera on facial recogni on
systems will, nevertheless, find extensive use in smart se ngs, where
computers and other devices serve more as helpful assistants.

Computers must be capable of accurately iden fying surrounding individuals in


a way that blends in with the natural rhythm of everyday human contact in
order to accomplish this goal. They must be adaptable to human percep ons of
when recogni on is likely and not necessitate specific interac ons. This
suggests that future intelligent environments should have roughly the same
modali es and constraints as humans. These objec ves currently seem to be
within reach, but there is s ll a significant amount of research to be done in
order to make human recogni on technology perform reliably in a wide range
of se ngs using data from a single or several modali es.

IIT GUWAHATI
23
REFERENCES

● Baron, R. J. (1981). Mechanisms of human facial recogni on.


Interna onal Journal of Man Machine Studies, 15:137-178.
● Heisele, B. and Poggio, T. (1999) Face Detec on. Ar ficial Intelligence
Laboratory. MIT.
● Frank, H. and Althoen, S. (1994). Sta s cs: Concepts and applica ons.
Cambridge University Press. P.110
● Brunelli, R. and Poggio, T. (1993), Face Recogni on: Features versus
Templates. IEEE Transac ons on Pa ern Analysis and Machine
Intelligence, 15(10):1042-1052.
● Figures are taken from internet.

IIT GUWAHATI

You might also like