2017 International Conference on Computational Intelligence and Networks

Artificial Intelligence techniques used to detect object and face in an image:

A Review

Deepika P.U Shivangi Chauhan Neetu Narayan

Deptt of Computer Science and Deptt of Computer Science and Deptt of Computer Science and
Engineering, ASET Engineering, ASET Engineering, ASET
Amity University Uttar Pradesh, Noida Amity University Uttar Pradesh, Noida Amity University Uttar Pradesh, Noida
Email ID: [email protected] Email ID: [email protected] Email ID: [email protected]

Abstract: In the modern world of digitalization, the need to without explicitly define the model mathematically. Here
develop expert system in growing tremendously. Most of the the assignment of each pixel is done with a probability
expert system perceive environment as image, and for reading value; then Skin Probability Map (SPM) is constructed
the components of image there are various techniques. This
using this value[1]. Dynamic skin distribution model is a
paper focuses on artificial intelligence algorithm that can be
used to extract features from the image. combination of all the other methods and is generally
executed in tracking applications, here the skin model is
Keywords: Artificial intelligence, Neural network, adjusted adaptively and automatically according to the
normalization, Object recognition, Template. training images as well as conditions dynamically.

I. INTRODUCTION B. Colorspaces used for skin modeling

Artificial Intelligence is a theory that is used to make A variety of color spacing techniques have been applied
expert system that can perform activities with the same in skin color modeling. Few of the color spaces and
intelligence as that of humans. Now-a-days use of their properties are as follows:
artificial intelligence to develop expert systems has been
increasing. Many expert systems require extraction of a) RGB
data from an image. Here in this paper different artificial RGB is a colorspace devised from CRT display
intelligence algorithms are dealt with that are used to applications, which is described in the form of three
extract data from the image. color rays (red, green and blue). The color-space is
mostly used for storing and processing of digital image
In this paper, different techniques have been reviewed data.
and discussed that extract the features from image which
therefore can be used for further data interpretation in b) Normalized RGB
other fields. Normalized RGB is a illustration which is achieved by
applying a normalization procedure on RGB values.
In the following section different methodologies are The values of the normalized components are known
discussed that will be used for the detection of the and their summation, wherein the third component
parameters such as face extraction, edge detection, holds no significance and therefore can be omitted,
object recognition. reducing the space dimensionality.

II. EXTRACTION OF FACE FROM c) HSI, HSV, HSL - Hue Saturation Intensity
THE IMAGE Colorspaces based on hue-saturation were developed
and used when the color properties were specified in the
Propose methodology present a three tiered hybrid face mathematic form. Hue defines the dominant color (for
model wherein skin color model is followed by edge example, yellow, red and purple) in the image data space
detection then face verification as shown in the figure. and saturation determines the colorfulness in the area of
the image data space.
A. Skin Modeling
d) YCbCr
According to the studies there are four methods of skin “YCbCr is an encoded nonlinear RGB signal, commonly
modeling; those are direct approach, parametric, non- used for image compression work. Color is represented
parametric and dynamic skin distribution models. Direct by luma, constructed as a weighted sum of the RGB
method helps in determining the intensity values that are values, and two color difference values Cr and Cb that
part of skin color explicitly. Parametric skin distribution are formed by subtracting luma from RGB red and blue
interpolates and generalizes the training data, and the components. The explicit separation and transformation
model is more representative and compact, as well as simplicity of luminance and chrominance components
independent from training data. Non parametric model makes this colorspace attractive.”
describes a skin model by assessing the training data

978-1-5386-2529-3/17 $31.00 © 2017 IEEE 6

DOI 10.1109/CINE.2017.20
III. EDGE DETECTION different expressions, lighting conditions, and
orientation. The implementation of the same is given
Input Image in the above Figure 2.

Color balancing and color spacing V. OBJECT RECOGNITION

using YCbCr method
Object Recognition is a technique of finding a
particular object in a given image. There are various
Skin color modeling techniques for object recognition some are given
below:
A. Template Matching
Edge detection using Canny edge
detection algorithm

Extraction and face from that of

background and also removal of
noise from the image

Face recognition and determination

of facial expressions using
Adaboost algorithm Fig.3. (a)Image Template (b)Image to be
Fig.1. Flowchart of the process of face detection compared with (c) Detected image superimposed
on grey-scale image[3]
Edge detection is performed to determine the
boundary line so as to differentiate the human face Template Matching is a technique in which two images
region from that of the background. There are various are used to find the object. It is used to find a small part
techniques to perform edge detection such as Canny, in an image by matching it with the template image.
SUSAN edge detection method etc. The Canny When an Image is input, it matches with the already
algorithm makes use of an optimal edge detector that stored Template image if the template image matches
is based upon number of standards that includes then it shows the presence of the object else that object
determining most of the edges by reducing the rate of is not there in the image. Hence in this way the object
error, maximizing the localization by marking edges present in the image is recognized.
close enough to that of the actual edges and by
marking up the edges once when there is single edge B. RECOGNITION BY COMPONENT(RBC)
present so as to minimize the response.[2] EDGE DETECTION
To find the edges of all the objects present in the image using any
edge detection technique.
IV. FACE DETECTION USING ADAPTIVE
BOOSTING ALGORITHM
Detection of non-accidental properties and parsing of concavities
This is done in order to see the viewing of the object and find at
what point geons are connected with each other.

DETERMINATION OF COMPONENTS
That is the number of geometrical shapes used and the relation
between them.

MATCHING OF COMPONENTS WITH OBJECT

REPRESENTATION
Now the components determined will be compared with the
representation given for an object.

OBJECT IDENTIFICATION
Fig.2. Process of Adaboost algorithm After comparing with object representation the final object is
identified.
AdaBoost, known as “Adaptive Boosting", is
a machine learning meta-algorithm. This algorithm Fig.4. Process of object recognition
detects human face in different scales, various poses,

7
 Recognition by component is a theory proposed by Table1. Comparison of Different Techniques
Irving Biederman. According to RBC theory we can used for Object Recognition
recognize objects present in a digital Image by matching
them with GEONS which are simple 3D geometrical Techniques Approach Result
shapes. According to Biederman there are 36 Geons in used for
total that are enough to recognize any kind of object. object
Two main components to be considered while looking recognition
for an object are: In this approach Template
1) Edges: Irrespective of the viewing angle of the object the matching of the matching is a
perception for the object remains the same. TEMPLATE input image is reliable
2) Concavities: this is the area where two geons meet. This MATCHING done with the technique.
helps us to recognize the boundaries of the geometrical given template and
shapes. The process for object recognition is illustrated recognize the
in the above Figure 4. object.
It is a flexible
C. NEURAL NETWORKS RECOGNITI In this objects are system.
ON BY divided into
Object recognition can be done by employing neural COMPONEN geometrical shapes
network[4]. The processing of a neural network System TS known as geons.
has been shown below: In this approach Artificial
Preprocessing: In this step we apply edge detection NEURAL through edge neural network
technique. Then we find edge pixels using USAN, and NETWORKS detection shape is is good at
then shapes are extracted from the image. extracted and handling
Descriptor creation: After extracting the shapes we through descriptors uncertanities.
define descriptors. There are two types of descriptors object is identified.
direct & invariant. Direct descriptor includes pixel Table2. Comparison of Different Techniques use
representation (represents a shape by its pixel) & for Color Spacing of an image
orientation descriptor (aims to decrease the quantity of
inputs that are been added from the pixel representation). Techniqu Approach Result
Radial representation comes under invariant descriptor es used
and uses circle to describe shape and try to extract as for color
much data as possible. spacing
Learning and recognition: Through various techniques RGB It helps in Helps in storing and
and the description about the shape we identify the processing the processing of digital
object present in the image. Hence a classified object is digital image and image data.
recognized. describe in the
form of three
VI. CONCLUSION colored rays
(red, green and
“In this paper, we have provided the description blue)
and comparison of most generally used methods for skin Normalize Achieved by the It is unchanged to
modeling and detection. We have summarized the most d RGB normalization of variations of surface
significant and notable differences between the method RGB values. orientation
which includes their approach, advantages and reasonably to the
disadvantages. The conclusions that are drawn are given light-source
below:” “HSI, Here the color It is unchanged to
HSV, HSL properties are the highlights at
- Hue defined white light sources,
Saturation numerically on as well as, “for
Intensity” the basis of tone matte surfaces, to
and saturation. ambient light and
surface orientation
relative to the light
source.”
YCbCr It is encoded Explicit separation
nonlinear RGB and transformation
signal. simplicity of
components makes
it popular.

8
 VII. FUTURE SCOPE [5] Irving Biederman “Recognition-by-Components: A
Theory of Human Image Understanding” in
These algorithms can be further used in various Psychological Review , Vol. M, No. 2,pp 115-
fields for feature detection in images. Also the 147,1917.
algorithms can be modified in such a way that the [6] Khushboo Khurana , Reetu Awasthi “Techniques
drawbacks of the technique could be overcome. for Object Recognition in Images and Multi-Object
Detection” in International Journal of Advanced
REFERENCES Research in Computer Engineering & Technology
(IJARCET) Volume 2, pp 1383-1388, April 2013.
[1] Setiawan Hadi et. al., Generating Skin Distribution [7] Lucy Bartosik “BIEDERMAN´S `RECOGNITION-
Map of Face Images, SNATI 2006, 17 Juni 2006 BY-COMPONENTS´” in 2002.
Yogyakarta [8] Herman Gomes “Marr's Theory: From primal sketch
[2] Jyoti Verma1, Vineet Richariya, “Face Detection and to 3-D models” in 2000.
Recognition Model based on Skin color and edge [9] Farzaneh Azadi pourghahestani, Esmat Rashedi
information for frontal face images.” “Object detection in images using artificial neural
[3] Kavita Ahuja, Preeti Tuli “Object Recognition by network and improved binary gravitational search
Template Matching Using Correlations and Phase algorithm” in 4th Iranian Joint Congress on Fuzzy
Angle Method” in International Journal of Advanced and Intelligent Systems 2015.
Research in Computer and Communication
Engineering , Vol. 2, pp 1368-1373, March 2013 .
[4] Jelmer de Vries “Object Recognition: A Shape-Based
Approach using Artificial Neural Networks” in 2006.