Site Loader
Rock Street, San Francisco

Aging Face Recognition System
Using Local Pattern Selections


Trupti Hankare

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!

order now

Shraddha Kasbe

UG Student


Department of
Computer Engineering

of Computer Engineering

Vidyapeeth College of Engineering

Vidyapeeth College of Engineering


Pranali Mhatre

Pratiksha Musale

UG Student


of Computer Engineering

of Computer Engineering

Vidyapeeth College of Engineering

Vidyapeeth College of Engineering


Prof.Madhuri Ghuge


Assistant  Professor

of Computer Engineering


Vidyapeeth College of Engineering




Aging is the biological process
of growing older in a deleterious sense. Automatic Face Recognition is based on
recognizing the faces of same person across different ages which has many real
time applications such as identifying missing children, identifying criminals.
The major cause of aging is facial features of same person is changed due to
aging process. Intra-user dissimilarity is used to describe the changes that
happened to the same person’s facial features due to aging. This paper is
proposed to solve the problem by introducing an effective method with
hierarchical learning model based on Local Pattern Selection (LPS). LPS
algorithm is used for low-level learning visual structures. It minimizes the
intra-user dissimilarity. The first level, low level features are extracted
using LPS descriptor and at the second level the output of first level is used
as an input to the second level to produce refined and accurate results.


Aging Face, Face Recognition, Intra-User Dissimilarity, Feature Descriptor.






Automatic face recognition is the biometric
identification by scanning a person’s face and matching it against a library of
known faces. Automatic age estimation has great interest because of its many
useful real-world applications-finding missing children, identifying criminals
based on identity ‘mug shots’ 5 9. Automatic face recognition is a
particularly attractive biometric approach, since it focuses on the same
identification that humans use primarily to distinguish one person from
another: their faces. One of its main goals is the understanding of the complex
human visual system and the knowledge of how humans represent faces in order to
discriminate different identities with high accuracy. Many methods of face
recognition have been proposed. Basically they can be divided into holistic
template matching based schemes, geometrical local feature based schemes and
hybrid schemes of all these types have been successfully applied to the task of
face recognition. Intra-user dissimilarity makes the task very difficult 1.


Automatic face recognition consists of subtask in a
sequential manner- face detection, face segmentation/face normalization and
face recognition/verification. By observing features of face, it is categorized
into three partitions – babies, younger adults, senior adults. First primary
features are extracted and then secondary features deals with wrinkle analysis that
helps in distinguishing senior adult from younger adult 6. The rest of this
paper is organized as follows. In section-II, a detailed explanation of system
is provided. In section-III, this paper presents an efficient feature
refinement framework. Section-IV introduces the simulation results. The paper
is concluded in the section-V.











rights reserved by






Face Recognition System Using Local Pattern Selections


(IJSTE/ Volume 2 / Issue 11 / 2018)


























Figure 1:
Example faces for one of the subjects in the MORPH Database





In this paper, a two-level
learning model is used to address the problem of intra-user dissimilarity 18.
In this system, effective features are learned from low-level microstructures
based on the feature descriptor called Local Pattern Selection (LPS). LPS
selects low level common information between cross-age faces. Then higher level
visual information is further refined based on output from first level.
Usually, in all image related projects, LBP feature extraction technique is
used. Local Binary Pattern (LBP) is simple and accurate but it maximizes the
intra-user dissimilarity where as LPS is used to minimize intra-user dissimilarity
16 17.




LPS is the feature extraction
method used here. Feature extraction is the technique of extracting features of
the image like its mean, variance, standard deviation, entropy etc. Now these
features stand instead of the image. 2


A.    Motivation


Facial image provides a pool of
information. The information regarding age, gender, identity, etc. Aging
estimation is the technique of estimating age from face. Aging occurs to
different person at different rates. Aging rate depends on the intrinsic
factors and extrinsic factor. Due to this persons of same age look quite
differently from one another. Intra user dissimilarity is the technical term
used here. Intra user dissimilarity is defined as the changes happening to the
same person during the process of aging. Sometimes due to the stressful working
condition intra user dissimilarity is very high because of that person at two
different age appears to be two different individuals. To overcome this
challenge the LPS is used. It reduces intra user dissimilarity and increases
common information thus increases the reliability of the system. The
implementation of feature extraction needs three steps-Pixel feature formation,
encoding tree and code assignment 2.


B.       Pixel feature formulation


Pixel features are the identity of the pixel. Pixel
feature is a 8 dimensional vector that is formed by comparing central pixel
with adjacent 8 neighbors taken at a radius r. The radius can take values r =
(1, 3, 5, 7). Here r is selected as 3. 18







rights reserved by






Face Recognition System Using Local Pattern Selections


(IJSTE/ Volume 2 / Issue 11 / 2018)






















Figure 2: Pixel feature


C.    Encoding Tree


Encoding tree is a binary tree
means zero or one. Here zero stands for NO and one stands for YES. The encoding
tree consist of leaf nodes and internal nodes. Leaf nodes are identified by
integers and internal nodes are identified by attributes and threshold




















.The pixel passes through the
encoding tree, it is either routed to left or right 18. Attributes are the
element of the pixel feature that is to be encoded. If the pixel meets the
threshold, it is routed to the right part otherwise to the left part. All the
pixel that falls to the same leaf nodes are provided with same codes.



Figure 3: Encoding tree with
internal and leaf nodes.




D.    Code Assignment


The pixels falls under the same
leaf nodes are provided with same code. For example, all the pixel that comes
under the leaf node 5 are assigned with the decimal code 5. 18 Thus minimizes
the intra user dissimilarity which is stated as the major challenge.





rights reserved by






Face Recognition System Using Local Pattern Selections


(IJSTE/ Volume 2 / Issue 11 / 2018)





The basic idea of this algorithm is to grow the
encoding tree incrementally until it reaches the expected number of leaf nodes


At each step, the node that maximizes the increase
in utility for expansion is selected.


The corresponding pixel features of these image
pairs are denoted as:


A = {(x-nm, y-nm)|m=1,…,M; n=1,…,N}



Where, M = H _W is the total number of pixels in an


We say x-nm and y-nm are matching pixel features as they belong to the same pixel location
in the images of the same subject 18. For any encoding tree T of L leaf
nodes, suppose the T assigns each pixel feature in A with a code based on which
leaf node that pixel finally reaches, resulting in an encoded set:


C= {(u-nm, u-nm)|m=1,…,M; n=1,…,N}


Let’s denote the tree with K leaf nodes (1 _ K < L) as   TK and the corresponding utility as UK. Now we extend TK into TK+1 by splitting node w into two children nodes wl and wr.   Denote  the sets of support pixel features of node w as:   S1w = (I1(1),….,In1(1) | Ii(1) €1,…,M*N} (3) S2w = (I1(2),….,In2(2) | Ii(2) €1,…,M*N} (4)     ALGORITHM 1: LOCAL PATTERNS SELECTION   Input: The number of leaf nodes L, the tradeoff factor _and training image pairs (I1n, I2n) n=1,…N} 18. Output: Encoding tree T.   /* Pixel features extraction. */   begin   Convert images into a set of pixel features as described in Eqn (1):   A  = {(x-nm, y-nm)|m=1,…,M; n=1,…,N}   /* Encoding tree initialization. */   begin   Initialize encoding tree T by adding one leaf node w, whose indices of support pixel features are:   S1w = (I1(1),….,In1(1) | Ii(1) €1,…,M*N} S2w = (I1(2),….,In2(2) | Ii(2) €1,…,M*N} w.a?0,w.t?0 and w.?u?-inf.   /* Encoding tree learning */   begin   for step = 2? L do   for each leaf node w do   if w.?u ? -inf then   /* Node has been evaluated. */   continue;   else   for k = 1?8 do   for z = min(S1w,S2w) ? max(S1w,S2w)do     All rights reserved by 4         Aging Face Recognition System Using Local Pattern Selections   (IJSTE/ Volume 2 / Issue 11 / 2018)     Evaluate increase of utility ?u with attribute = k and threshold = z.   Let maximum ?u* is achieved at (k*; z*).   Update: w.?u = ?u*, w.a ?k*, w.t? z*.   Let w*has maximum _u over all leaf nodes.   Split w* into two children nodes l and r.   l.a ?0, l:t   0, and l.?u ? -inf.   r.a ?0, r:t  0, and r.?u ? -inf.   Update S1l ,S2l, S1r ,S2r based on Eqn (3),(4).   Assign distinct codes to leaf nodes, and return T. 18     IV. LPS BASED FEATURE EXTRACTION   The Algorithm 1 learns an encoding tree that encodes given image by converting each pixel into decimal codes based on the leaf node that pixel reaches in the tree. In this part, we briefly introduce how we extract over-completed features based on encoded images. The techniques we use include multiple scaling and dense sampling 8. Specifically, we first train multiple encoding trees based on different sampling radii (e.g.1; 3; 5; 7) as illustrated in Figure 2. Then for each encoded image, we extract local features by calculating the histograms of small patches formed by dividing image into overlapping (with overlapping factor 0:5) fixed size (e.g. 16 _ 16) areas. The final features of an image are formed by concatenating local features at all sampling radii.   V. INFORMATION REFINEMENT   Feature extraction technique LPS is used to collect low level visual structures. The features collected using LPS are of high dimension due to the employment of multiple scaling and dense sampling technique. The high dimension features arises problems in storage and matching. Same features may be there for many times. Redundant data and noise may mislead to wrong decision. So it is essential to filter out these redundant data and noise, thus making set of features more precise. Filtering points to the need of refinement technique 2. The most popular refinement technique commonly used in image processing is Fisher linear discriminant analysis or universal subspace analysis (USA). It deals with the eigen decomposition of a matrix of size Min (N, D) where N is the number of training image pairs and D is the dimension of features. The time complexity of eigen decomposition is around min {O (N3), O (D3)}.Thus making it not suitable for very large values of N and D 2 18. Here this part presents a refinement technique that successfully eliminates redundant data and noise, thus provides crucial data for the classification. This refinement technique has two stages. First one is the bootstrap aggregating and the second is random subspace classifiers.   A.   Bootstrap aggregating: Divide Data by Sampling   Bootstrap aggregating is otherwise known as bagging. The one line definition of the technique is dividing data by sampling. The data set include pictures of 10,000 persons that means N = 10,000, N is defined as the no: of training image pairs. In the data set two photographs of same person taken at a large age gap are included. . Bagging technique is used to select a series of subset M (M

Post Author: admin


I'm Dora!

Would you like to get a custom essay? How about receiving a customized one?

Check it out