TAG ME: An Accurate Name Tagging System for Web Facial Images using Search-Based Face Annotationv

International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
753
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
TAG ME: An Accurate Name Tagging System for Web Facial Images using
Search-Based Face Annotation
Mr. Ansari Mohammed Abdul Qadir Ataullah
Department of Computer Engineering
MET’S Institute of Engineering, Bhujbal Knowledge City
Adgaon Nashik, University of Pune, Maharashtra, India
abdul11q@gmail.com
Prof. R.P.Dahake
Department of Computer Engineering
MET’S Institute of Engineering, Bhujbal Knowledge City
Adgaon Nashik, University of Pune, Maharashtra, India
dahakeranjana@gmail.com
Abstract— Now a day the demand of social media is increases rapidly and most of the part of social media is made up of multimedia content
cognate as images, audio, video. Hence for taking this as a motivation we have proffer a framework for Name tagging or labeling For Web
Facial Images, which are easily obtainable on the internet. TAG ME system does that name tagging by utilizing search-based face annotation
(SBFA). Here we are going to select an image from a database which are weakly labeled on the internet and the "TAG ME" assign a correct and
accurate names or tags to that facial image, for doing this a few challenges have to be faced the One exigent difficulty for search-based face
annotation strategy is how to effectually conduct annotation by utilizing the list of nearly all identical face images and its labels which is weak
that are habitually rowdy and deficient. In TAGME we have resolve this problem by utilizing an effectual semi supervised label refinement
(SSLR) method for purify the labels of web and nonweb facial images with the help of machine learning techniques. Secondly we used convex
optimization techniques to resolve learning problem and used effectual optimization algorithms to resolve the learning task which is based on the
large scale integration productively. For additionally quicken the given system, finally TAGME system proposed clustering-based
approximation algorithm which boost the scalability considerably.
Keywords—Face annotation, SBFA, machine learning, semi supervised label refinement, web and nonweb facial images, weak label
__________________________________________________*****_________________________________________________
I. INTRODUCTION
As we know now a day a rapid growth of social
media increases day by day Due to that photo sharing and
tagging is very popular. As we see on every social media Most
of the content is established on images and images act as a one
of the big entertainment Part of social media. Everybody
wants to share their photos, images with each other on Social
media sites and on World Wide Web. Contemporary years
have witnessed a detonation of the Number of digital photos
taken and keep by consumers. An extra piece of photos Shared
by users online on social media are face images of human.
Few of these face images are tagged properly with proper
names, but numerous of them are improperly tagged. This
notion motivated the study of auto face annotation, which is a
dominant technique that point to annotate facial images
automatically [14].
All of the facts discuss above we can say that a "TAG
ME" system is advantageous to numerous actuality
applications, For example, by utilizing auto face annotation
techniques, online social networking sites which supports
sharing of photos (e.g., Facebook, twitter etc.) can self-acting
annotate user’s uploaded photos to make easier online photo
search and administration. Apart from this skill can be used in
news domain and in video domain to notice main persons
become visible in the videos to make easier the retrieval and
characterization task from news video. Classical methods of
annotation of face image are continually act towards an
enlarge face recognition problem. Nevertheless, the Model
based face annotation techniques are few within several facets.
First, it is habitually time consuming and costly to gather a
huge amount of training images of human faces which is a
labeled images. Second, habitually it is hard to generalize the
models when new data which is trained or new persons are
added, in which a thorough process is normally required.
Lastly, the recognition/annotation performance habitually plate
badly when the number of classes/persons is very large.
Currently, a few appearing studies have try to traverse a
encouraging search-based annotation concept for the
annotation of face images by mining online ,offline and
realtime facial images where a large number of facial images
which is weakly labeled are freely obtainable. As a substitute
of training external classification models by the frequent
model based annotation of face approach, the search-based
face annotation (SBFA) model point to tackle the automated
annotation of face work by utilizing CBIR i.e (content based
image retrieval) scheme within mining gigantic face images
which are weakly labeled online. The SBFA approach is data-
driven and model-free, which too little size is inspired by the
search based image annotation techniques for collective image
annotations [16].

Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
754
_______________________________________________________________________________________
II. RELATED WORK
This System is closely homologous to several
categories of research work and divided into five categories;
now let us see these categories one after another in details:
The first category of alike task is basically focus on the
concept of recognition and verification of faces, and it is old
research issues of pattern recognition and computer vision and
it studied from several years it is discuss in [03], [04]. Now
currently it is notice that a few appearing touchstone studies of
the detection and verification of face techniques on face
images are collected from various sources, cognate as the
LFW standard studies [05], [06], [07].
The second category is homologous with the studies
of common image annotation i.e frequently used annotation.
The old image annotation approaches habitually used few live
object recognition techniques for training a classification
models from the training images which are labeled by the
human being and try to deduce the probabilities/correlation in
the middle of annotated keywords and images. Giving short
trained data, semi supervised methods of learning may be
utilize for annotation of image [08]. This scenario is described
in Wang et al. [09] and Pham et al. [10] both of them put
forward the technique to purify the results based on model
based annotation with a label likeliness graph which follows
random walk principle. These all problems of image
annotation and its different solutions are discussed in Likely,
Pham et al. [10] and Russell et al. [11]. Unalike these live
works of the different peoples, TAGME put forward a semi
supervised label purification/refinement strategy which
concentrates firstly on optimizing the label quality for face
images towards the search-based face annotation task.
The third category focuses on annotation of a facial
image on various kinds of photos such as family
photos/personal photos and social photos. A few learning’s
[12],[14] have basically concentrate on the work of annotation
on different photos mainly family photos which frequently
include rich contextual clues, cognate as family/personal
names, social media surroundings, geotags, timestamps along
with others. The number of classes/persons is widespread
utterly compact, creating cognate annotation tasks not so much
exigent. All of these schemes widespread attain justly error
free annotation results, out of this some techniques have been
flourishing exploit in many applications which is
commercially used, for example, Picasa by Google, Photo by
Apple, EasyAlbum by Microsoft [13], and the face auto
tagging solution of Face book.
The fourth category is regarding the learning’s of
annotation of face images in mining facial images which are
weakly labeled present on internet. Few learning’s consider
input query as a human name, and mostly point to purify the
text-based search outcomes by utilizing visual constancy of
face images. For example, Ozkan and Duygulu [15] put
forward a graph-based model for piercing the densest subgraph
as a lot alike solution. Utilizing the graph-based scheme, Le
and Satoh [16] put forward a contemporary native density
score which state the significance of all choose facial images,
secondly the Guillaumin et al. [17] put forward a moderation
which absorb the restraint i.e a face is at most portray ever
inside an image. Similarly, on the other side the productive
scheme equivalent to the model called as gaussian mixture
model which additionally chooses to the name-based search
strategy [02] and attain proportional solution. Freshly, one of
the distinguished concepts was proffer within [18] to enhance
over the productive secheme. Utilizing plan from expansion of
query [19], the production of name based strategy may
additionally enhance with inaugurating the friends images as
the name of query. Unalike these learning are of sieving the
solution based on text-based retrieval, a few learning’s have
Endeavour to absolutely annotate all face images with the
specific names comes from the caption information, Berg et al.
[20] put forward a likelihood technique and also include the
clustering algorithm to calculate the names of the caption and
their facial images. In TAGME the task is dissimilar from the
above precursory tasks in terms of two main facets. Firstly,
TAGME resolve the widespread content-based face annotation
problem utilizing search-based archetype, here images of face
are absolutely consider as a query images and the work is to
give correlating names of the images taken as a query. Only
few learning’s has been described on this kind of idea. A few
contemporary works [21] mostly tackle the face retrieval
problem, within which is an effectual image portrayal has been
proffer utilizing both global besides local features. Secondly, it
established on beginning labels which are weak, the proffer
semi supervised label purification/refinement algorithm
acquire knowledge of an amplified new label matrix having
whole face images within the total name space.
The fifth category concerning the learning’s of
cleansing face images, which point to leverage rowdy online
facial images for face recognition supplications. Habitually
this kind of works have proffer as an easier preprocessing pace
within the entire system without using worldly schemes.
Consider the e.g., the task in [02] put in a modified k-means
clustering approach for cleansing up the rowdy online face
images. Zhao et al. [22] have proffer a constancy method of
learning for face models training of the celebrity with the help
of mining the text-image co-occurrence on the internet as a
poor signal which is applicable in the direction of supervised
learning task of faces from a bigger and rowdy training set.
Unalike the given live tasks, the system utilize the semi
supervised machine learning schemes and put forward a label
purification/refinement algorithm which is graph based to
enhance standards of the image label over the entire retrieval
database in the SBFA task [01].

Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
755
_______________________________________________________________________________________
III. PROPOSED SYSTEM
TAGME system flow architecture is shown in fig. 1
the overall system represented in three modules:
a) Facial Image Gathering.
b) Peprocessing, Learning & Indexing.
c) Retrieval and Image Annotation.
Fig. 1. TAGME System Flow Architecture
The short descriptions of these modules are as follows:
(a) Gather labeled facial images from internet by utilizing web
search engines.
(b) Preprocess the facial images, By Implementing face
detection, feature extraction and alignment for the detected
faces; following that, TAGME Used indexing to extracted
high-dimensional facial features and used proffer SSLR
scheme to purify the labels simultaneously with the proffer
clustering-based approximation algorithms for improving the
scalability.
(c) Search for the query facial image to retrieve the top K alike
images and use their analogous names for voting toward auto
annotation.
The above paces are broadly classified as:
1. Facial image data gathering.
2. Face feature extraction and its detection.
3. High-dimensional feature indexing.
4. Indexing and learning for purification of weakly
labeled data.
5. Alike face retrieval.
6. Annotation of face by percentage of matching on the
alike faces with the purified labels.
The first four paces are habitually conducted before
the face annotation task test phase and remaining paces face of
annotation task are conducted in between the test phase, which
habitually done accurately. We describe these paces one by
one below.
The first pace dealing with the data gathering of
facial images as seen in Figure 1, in which system gathering
facial images from the Internet using a live web search engine
(i.e., Google) as per the searched query given. The output of
this crawling process shows the facial images of a given
person and each of them is analogous with a few human
names. As per the nature of web images, these facial images
are continually rowdy, which do not consistently related to the
correct human name. Thus, the system calls cognate category
of facial images with rowdy names as weakly labeled facial
image data.
The second pace is to preprocess these facial images
taken from online to extract the information homologous with
the human face, this preprocessing includes including
detection of face, its alignment, extraction of facial region
extraction and facial feature portrayal. For the detection of
face and alignment, the system embraces the semi supervised
face alignment technique proffer in [23]. For facial feature
portrayal, the system extracts the GIST texture features to
depict the extracted faces. The result of these processes is that,
each face can be depicted by a d-dimensional feature vector.
The third pace is indexing the features which are
extracted from the face by implementing a few efficient high-
dimensional hash indexing techniques to facilitate the retrieval
of similar face task. In this approach, system choose the
locality sensitive hashing (LSH), a very famous and effectual
hash based indexing technique. Apart from the indexing pace,
an additional vital pace of TAGME is to occupy a semi
supervised learning strategy to intensify label standards of
facial images which is weakly labeled. This is very salient
process for the whole framework of search based annotation,
after all the label standards act as a judge mental factor in the
final annotation performance.
Each of the above is the activities prior to the
annotation of a query facial image. Next, now the system
reports the activity of annotation of face throughout the test
phase. In specific, considering a query facial image for
annotation, the system firstly supervised alike retrieval of face
process for finding a subset of most alike faces (generally top
K alike faces) from the precursory dataset consist of facial
images which is indexed using hash indexing. Along the set of
top K alike faces which is retrieved from the dataset, the pace
is the face annotation with a label by utilizing a percentage of
matching approach that merge the labels analogous with these
top matching alike faces. In this topic, the system focuses our
attention on one key pace of the above framework, i.e., the
semi supervised process of learning to purify labels of the
labeled facial images.
A. Algorithms
1) Multistep Gradient Algorithm Utilizing SSLR
The optimization tasks belong to exactly quadratic
programming (QP) problems. It appears to be feasible to
resolve them absolutely by applying generic QP resolvers.
Nevertheless, this would be computationally extremely
thorough after all matrix F can be probably huge, for example,
for a big 400-person dataset of entirely 40,000 face images, F
is a 40,000* 400 matrix that resides of 16 million variables,
which is almost infeasible to be resolved by any live generic
QP resolver. So we first select multistep gradient algorithm to
resolve the problem, as shown in algorithm.

Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
756
_______________________________________________________________________________________
Parameters Meaning
Q Є R(n.m)x(n.m)
Vectorizing Matrix
x*
Optimal Solution
K No. of Iteration
Z(k)
Search Point for Label Refinement
t Lipshitz Constant
α0 Regularization Parameter, consistently α>0.
Table 1. Parameters Used in Algorithm.
Algorithm :
Input : Vectorizing Matrix Q Є R(n.m)x(n.m)
Output : X* i.e Optimal Solution
Start
Step 1 : Initialize Parameters:
1. Regularization Parameter (α0) Initially Set to
One
2. No. of Iteration (k) Initialize to One
3. Search Point (Z(k)
) is Initially Set to Zero
4. Optimal Solution at Start (X0
) is Set to Zero
5. Optimal Solution Minus One (X-1
) is Set to Zero
Step 2 : Repeat Following Steps Until Convergence
1. Used Soft Regularization Formulation to Attain
Approximate Solution (Xk
)
2. Used Convex Constrain Formulation to Achieve
)
3. Calculate Regularization Parameter(αk)
4. Process the (Z(k)
) for Combining Two Previous
) and (Xk-1
)
5. Increase the Value of Interation i.e K = K+1
Stop
B. Mathematical Model
Let the TAGME system is described by S,
S = (IC, FD, FI, LD, FR, FA)
Q = (q1,q2,.qn)
Ω = (n1,n2,n3.....)
T = (t1,t2,t3......)
R = Dataset
X*
= Optimal Solution
Where
S : Depicts TAGME System.
Q : Set of Input Query Images.
R : Dataset.
Ω : Human Names List for Annotation.
T : No. of Iterations to find Enhanced Image.
IC : Gathered the data of facial images
FD: Detected Face and Extracted Features.
FI : Indexing of Features of Face.
LD: Data Learning and Purification of Labels.
FR: Retrieval of Facial Images
FA: Annotation of Face Using Percentage of
Matching
X*
: Annotated Image.
Fig. 2. Vein Diagram of TAGME
IV. IMPLEMENTATION AND RESULTS
A. Performance Evaluation of Face Detection.
In this experiment we have calculated the face detection
performance on a celebrity images which is used in TAGME
system database, the details of this operation is as shown in
table 2.
Sr.
No.
Celebrity Name
&
Its Dataset Name
Total
Images
Taken
Detected
Images
Recognized
Images
Offline Images
1 Image_set1 15 13 10
2 Imageset2 15 10 08
3 Image_set3 15 10 09
4 Image_set4 15 14 12
5 Image_set5 15 11 08
6 Image_set6 15 13 10
7 Image_set7 15 14 13
8 Image_set8 15 14 12
9 Image_set9 15 13 11
10 Image_set10 15 13 12
Online Images
1 Image_set11 16 10 08
2 Image_set12 14 08 05
3 Image_set13 14 09 07
4 Image_set14 13 08 07
5 Image_set15 16 10 08
6 Image_set16 12 07 05
7 Image_set17 14 08 07
8 Image_set18 13 08 06
9 Image_set19 16 09 07
10 Image_set20 15 08 05
Realtime Images
1 Image_set21 10 06 03
2 Image_set22 07 04 02
3 Image_set23 12 06 04
Table 2. Face Detection Performance

Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
757
_______________________________________________________________________________________
Table 2 is depicted utilizing graph as shown in fig.3, here we
consider the part of the overall dataset and represented as
Image_set of different celebrities, out of the total images
which is taken from the Image_set we have find the detected
faces and recognized faces.
Fig. 3. Face Detection Performance Graph of Offline Images
B. Evaluation of Image Clustering
In this experiment we shows overall procedure of image
clustering in this process firstly dissimilar clusters of a
identical image are formed as per the percentage of matching,
for example if we give shahrukh khan image for retrieval then
all the shahrukh khan images from the database are shown in
fig. 4, following that we perform clustering on that image and
then shows clustering results according to the percentage of
matching, means exact alike image shows 100% alike and
remaining images are display as per their matching percentage,
this is possible by creating a dissimilar clusters of a identical
image and match that cluster with the input image cluster. The
process of clustering is depicted as follows, here
C1,C2,C3…..C7 depict the dissimilar cluster of a identical
person according to the percentage of matching. Hence
utilizing clustering we calculate the annotation performance of
TAGME system.
Fig. 4. Evaluation of Image Clustering
C. Evaluation of Annotation
This experiment points to examine the automatic annotation of
face established using the search based face annotation. In
TAGME the annotation is done by utilizing percentage of
matching with the images in the database, we evaluate this on
a database consists of 3000 images. Figure 5 shows the
annotation count with respect to the number of images of a
celebrity present in a database, here we consider total images
of different celebrities from a database and shows match
annotation from the database of that celebrity in percentage
form.
Fig. 5. Evaluation of Annotation
V. APPLICATIONS
1) Over Internet For Accurate Name Tagging Of Facial
Images.
2) Over Social Media Sites.
3) Over Intranet Of an Organization.
4) In a Biometric Security System For Giving Name to the
Facial Image.
VI. CONCLUSION
"TAG ME" mostly focused real life problem of name
tagging over social media, over the internet and additionally
on tackle the sever problem to enhanced the label standards,
Secondly this system could be used for upgrade in the
scalability and for successful acceleration of the task
optimization without any degradation of performance. Here we
additionally introduced performance annotation analysis, real-
time images as an input, work on group images also the
annotation criteria is set as per the percentage of matching and
we also consider variable image sizes. Hence, we can say that
the problems of existing system w.r.to name tagging over
internet is remove here, and the ability of searching the correct
image according to the input label or name is increasing in
TAG ME and it is also pertinent for large scale database.
0
2
4
6
8
10
12
14
16
No.ofImages
Image Set of Different Celebrity
Total Images Detected Recognized Images
0
10
20
30
40
50
60
No.ofImagesofaCelebrity
Image Set of Different Celebrity
Total Images Annotation Count

Volume: 5 Issue: 7 753 – 758
_______________________________________________________________________________________________
758
_______________________________________________________________________________________
ACKNOWLEDGEMENT
We are very much thankful to all the authors of the papers
which we have referred and glad to express my sentiments of
gratitude to all of them who rendered their valuable
contribution and help to make this work successful.
REFERENCES
[1] Dayong Wang, Steven C.H. Hoi,Ying He,Jianke Zhu, Mining
Weakly Labeled Web Facial Images For Searched-Based Face
Annotation, IEEE Transaction on Knowledge and Data
Engineering, Vol-26 2014.
[2] T.L. Berg, A.C. Berg, J. Edwards, M. Maire, R. White, Y.W.
Teh, E.G. Learned-Miller, and D.A. Forsyth, Names and
Faces in the News," Proc. IEEE CS Conf. Computer Vision
and Pattern Recognition (CVPR), pp. 848-854, 2004.
[3] P. Belhumeur, J. Hespanha, and D. Kriegman, Eigenfaces
versus Fisherfaces: Recognition Using Class Specific Linear
Projection," IEEE Pattern Analysis and Machine Intelligence,
vol. 19, no. 7,pp. 711-720, July 1997.
[4] W. Zhao, R. Chellappa, P.J. Phillips, and A. Rosenfeld, Face
Recognition: A Literature Survey," ACM Computing Survey,
vol. 35, pp. 399-458, 2003.
[5] G.B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller,
Labeled Faces in the Wild: A Database for Studying Face
Recognition in Unconstrained Environments, technical report
07-49, 2007.
[6] H.V. Nguyen and L. Bai, Cosine Alikeity Metric Learning for
Face Verification, Proc. 10th Asian Conf. Computer Vision
(ACCV 10), 2008.
[7] M. Guillaumin, J. Verbeek, and C. Schmid, Is that You?
Metric Learning Approaches for Face Identification, Proc.
IEEE 12th Intl Conf. Computer Vision (ICCV), 2009.
[8] Z. Cao, Q. Yin, X. Tang, and J. Sun, Face Recognition with
Learning Based Descriptor, IEEE Conf. Computer Vision and
Pattern Recognition (CVPR), pp. 2707-2714, 2010.
[9] C. Wang, F. Jing, L. Zhang, and H.-J. Zhang, Image
Annotation Refinement Using Random Walk with Restarts,
14th Ann. ACM Intl Conf. Multimedia, pp. 647-650, 2006.
[10] P. Pham, M.-F. Moens, and T. Tuytelaars, Naming Persons in
News Video with Label Propagation, Proc. VCIDS, pp. 1528-
1533, 2010.
[11] B.C. Russell, A. Torralba, K.P. Murphy, and W.T. Freeman,
LabelMe: A Database and Web-Based Tool for Image
Annotation, Intl J. Computer Vision, vol. 77, nos. 1-3, pp.
157-173, 2008.
[12] Y. Tian, W. Liu, R. Xiao, F. Wen, and X. Tang, A Face
Annotation Framework with Partial Clustering and Interactive
Labeling, Proc. IEEE Conf. Computer Vision and Pattern
Recognition (CVPR), 2007.
[13] J. Cui, F. Wen, R. Xiao, Y. Tian, and X. Tang, EasyAlbum:
An Interactive Photo Annotation System Based on Face
Clustering and Re-Ranking, Proc. SIGCHI Conf. Human
Factors in Computing Systems (CHI), pp. 367-376, 2007.
[14] J.Y. Choi, W.D. Neve, K.N. Plataniotis, and Y.M. Ro,
Collaborative Face Recognition for Improved Face
Annotation in Personal Photo Gatherings Shared on Online
Social Networks, IEEE Trans. Multimedia, vol. 13, no. 1, pp.
14-28, Feb. 2011.
[15] D. Ozkan and P. Duygulu, A Graph Based Approach for
Naming Faces in News Photos, Proc. IEEE CS Conf.
Computer Vision and Pattern Recognition (CVPR), pp. 1477-
1482, 2006.
[16] D.-D. Le and S. Satoh, Unsupervised Face Annotation by
Mining the Web, Proc. IEEE Eighth Intl Conf. Data Mining
(ICDM), pp. 383-392, 2008.
[17] M. Guillaumin, T. Mensink, J. Verbeek, and C. Schmid,
Automatic Face Naming with Caption-Based Supervision,
Proc. IEEE Conf. Computer Vision and Pattern Recognition
(CVPR), 2008.
[18] M. Guillaumin, T. Mensink, J. Verbeek, and C. Schmid, Face
Recognition from Caption-Based Supervision, Intl J.
Computer Vision, vol. 96, pp. 64-82, 2011.
[19] T. Mensink and J.J. Verbeek, Improving People Search Using
Query Expansions, Proc. 10th European Conf. Computer
Vision (ECCV), vol. 2, pp. 86-99, 2008.
[20] T.L. Berg, A.C. Berg, J. Edwards, and D. Forsyth, Who’s in
the Picture, Proc. Neural Information Processing Systems
Conf. (NIPS), 2005.
[21] Z. Wu, Q. Ke, J. Sun, and H.-Y. Shum, Scalable Face Image
Retrieval with Identity Based Quantization and Multi-
Reference Re-Ranking, Proc. IEEE Conf. Computer Vision
and Pattern Recognition (CVPR), pp. 3469-3476, 2010.
[22] M. Zhao, J. Yagnik, H. Adam, and D. Bau, Large Scale
Learning and Recognition of Faces in Web Videos, Proc.
IEEE Eighth Intl Conf. Automatic Face and Gesture
Recognition (FG), pp. 1-7, 2008.
[23] J. Zhu, S.C.H. Hoi, and L.V. Gool, Unsupervised Face
Alignment by Robust Non rigid Mapping, Proc. 12th Intl
Conf. Computer Vision (ICCV), 2009.

TAG ME: An Accurate Name Tagging System for Web Facial Images using Search-Based Face Annotationv

More Related Content

Similar to TAG ME: An Accurate Name Tagging System for Web Facial Images using Search-Based Face Annotationv (20)

More from rahulmonikasharma (20)

Recently uploaded (20)

TAG ME: An Accurate Name Tagging System for Web Facial Images using Search-Based Face Annotationv