Efficient and Robust Detection of Duplicate Videos in a Database

International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 11 139 – 142
_______________________________________________________________________________________________
139
IJRITCC | November 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
Efficient and Robust Detection of Duplicate Videos in a Database
Ragho Soni R.
Department of Computer Science and Engineering
M. S. Bidve Engineering College, Latur
e-mail: sonirragho@gmail.com
Shah H.P.
Department of Computer Science and Engineering
M. S. Bidve Engineering College, Latur
e-mail: shahhemali@gmail.com
Abstract— In this paper, the duplicate detection method is to retrieve the best matching model video for a given query video using fingerprint.
We have used the Color Layout Descriptor method and Opponent Color Space to extract feature from frame and perform k-means based
clustering to generate fingerprints which are further encoded by Vector Quantization. The model-to-query video distance is computed using a
new distance measure to find the similarity. To perform efficient search coarse-to-fine matching scheme is used to retrieve best match. We
perform experiments on query videos and real time video with an average duration of 60 sec; the duplicate video is detected with high similarity.
Keywords-Video fingerprintingt; color layout descriptor; distance mesure; vector quantization.
__________________________________________________*****_________________________________________________
I. INTRODUCTION
With the fast development of technology and increasing
use of the widespread accessibility of ADSL and the World
Wide Web, people can easily find and upload tons of videos
on the internet. There exist too many duplicated and
transformed video clips online and some of them may be
illegally copied or broadcasted, so database and copyright
management have become big issues nowadays.
Copyright infringements and data piracy have recently
become serious concerns for the ever increasing online video
database management. Videos on commercial sites e.g.,
www.google.com, www.YouTube.com, www.metacafe.com,
are mainly textually tagged. These tags are of little help in
monitoring the content and preventing illegal upload. There
are two approaches to detect such infringements that are
watermarking and Content-Based Copy Detection (CBCD).
The watermarking approach tests inserting distinct pattern in
video to decide if it is copyrighted. The other approach CBCD
detects the duplicate by comparing the fingerprint of the query
video with the fingerprint of the original video.
A video ―fingerprint‖ is a feature extracted from the video
that should represent the video compactly, allowing faster
search without compromising the retrieval accuracy. The
bottleneck of watermarking is that the inserted marks are
likely to be destroyed or distorted as the format of the video
get transformed or during the transmission so noise robustness
of the watermarking schemes is not ensured in general [1],
where as the video fingerprint extraction of the Content-Based
Copy Detection can be mostly unchanged after various noise
attacks, that’s why video fingerprinting approach has been
more successful.
Duplicate video is derived from only one database video it
consists entirely of a subset of frames in the original video.
The individual frames may be further processed. The temporal
order of the frames may also be altered. In [2], a set of 24
queries searched in YouTube, Google video and yahoo video,
27% of the returned relevant videos are duplicates or nearly
duplicate to the most popular version of video in the search
result. In [3], each web video in the database is reported to
have an average of five similar copies. Also, for some popular
queries to the yahoo video search engine, there were two or
three duplicates among the top ten retrievals [4].
II. RELATED WORK
Feature representation: In the feature extraction process,
feature extracted from the video and image fall into three
categories as global image, keypoint based, and Entire Video
based Features. Many technique use global features for a fast
initial search to find duplicates using color histogram
computed over the entire video [2] for coarse search and
keyframe-based features are use for a more refined search. A
comparative study for video copy detection methods can be
found in [5].
Global Image Features are derived from sets of time-
sequential video keyframes. A combination of MPEG-7
features such as the Scalable Color Descriptor, Color Layout
Descriptor (CLD) [6] and the Edge Histogram Descriptor
(EHD) has been used for video-clip matching [7], using a
string-edit distance measure.
Keypoint based feature technique described in [8] by Joly,
this includes a key-frame detection, an interest point detection
in these key-frames, and the computation of local differential
descriptors around each interest point. Interest points are ideal
for matching applications because of their local uniqueness
and their high information content. In [9]The SIFT descriptors
by Lowe use the Divergence of Gaussian (DoG) interest point
detector which better handles significant changes in the scale
of images. The SIFT descriptors then encode the image
gradients and their orientations around the points into a 128-
dimensional histogram. The PCA-SIFT descriptors simply
apply PCA on Lowe's SIFT descriptors, reducing their
dimensionality to 36. From the SIFT family this scheme is
called EFF2
as these descriptors are computed EFFiciently and
yield EFFective search results. Overall, local descriptor
schemes handle rotations, translations of objects in images,
changes in color and to some extent compression and scale
changes.
Entire video based features derived from the whole video
sequence, such descriptors like ordinal measure, have poor
performance with local transformations such as shift, cropping
and cam coding.
Indexing method: a number of indexing techniques have
been used for speed up the detection process .In [10] For the
videntifier system the NV-Tree [9], which is able to perform
accurate approximate High-dimensional nearest neighbor

Volume: 5 Issue: 11 139 – 142
_______________________________________________________________________________________________
140
_______________________________________________________________________________________
queries in constant time, independent size of the video
descriptor collection. In [11] author evaluate approximate
search paradigm, called Statistical Similarity Search (S3) in a
complete CBCD scheme based on video local fingerprints.
The proposed statistical query paradigm relies on the
distribution of the relevant similar fingerprints. Joly [8] use an
indexing method based on the Hilbert’s space filling curve
principle and a simple search method is used to find closest
derived key in the database.
Hash-based Index: Locality Sensitive Hashing (LSH)
[12], have been effectively useful for similarity indexing in
vector spaces and string spaces under the Hamming distance (a
well-liked approximate for L2 distances and in this proposed
distance function is non-metric). LSH formalism is not
applicable for analyzing the behavior of these tables as index
structures DBH is a hash-based indexing method that is
distance based. Consequently, DBH can be applied in
arbitrary (and not necessarily metric) spaces and distance
measures, Whereas LSH cannot.
Final Duplicate Confirmation: From the top ten Nearest
Neighbors, the system is supposed to return answer whether or
not it is a duplicate of an already existing copyrighted database
video. In[5][13],the partial result must be post-processed to
compute a similarity measure and to decide if the more similar
video is copy of copyrighted video using voting strategy. A
robust voting algorithm utilizes trajectory information, spatio
temporal registration, and label of behavior indexing to make a
final decision.
III. IMPLEMENTATION DETAILS
A. System Architecture
In Fig. 1, denotes system architecture of duplicate video
detection. This system works in two phases, offline phase and
online phase. Offline phase (model related) consists of the un-
quantized model fingerprint generation, VQ design and
encoding of the model signatures, and computation and storing
of appropriate distance matrices. Online phase (query related)
consist query video is decoded, sub-sampled, key frames are
identified, and features are computed per keyframe. It obtain
k-means based compact signatures, perform VQ-based
encoding on the signatures to obtain sparse histogram-based
representations, compute the relevant lookup tables, and then
perform two-stage search to return the best matched model
video.
In this paper the database video referred as original or
model videos. Model video is decoded and sub-sampled at a
factor of number to get (T) frames and feature (P) is extracted
from per frame. To summarize feature, we perform k-means
clustering and save the cluster centroid as fingerprint (F).The
duplicate video detection task return best matching model
fingerprint for query fingerprint. The model video to query
video distance is computed using distance measure. Two phase
procedure is used for search; coarse search is used for to return
top K Nearest Neighbors, refine search returns best match for
given query video. The final module decides whether the
return video is duplicate video or not.
Figure 1. System Architecture
B. Feature Extraction
1) CLD: The video is decoded frames are sub-sampled at a
factor of value to obtain frames and a vector dimensional
feature is extracted per frame using CLD technique. The CLD
is very compact and resolution invariant representation of
color for high speed image retrieval and it has been designed
to efficiently represent spatial distribution of color. The
extraction process of this color descriptor is obtained by
converting the frame to 8*8 into 64 blocks to guarantee the
invariance to resolution or scale. After the image partitioning
stage, a single representative color is selected from each block,
on averaging, along each(Y/Cb/Cr) channel. Y/Cb/Cr is
transformed by 8x8 DCT, so three sets of 64 DCT coefficients
are obtained. A zigzag scanning is performed with these three
sets of 64 DCT coefficients, The DC and first few AC DCT
coefficients for each channel constitute the CLD feature
dimension. Color Layout Descriptor (CLD) [6] achieved
higher detection accuracy than other candidate features. To
summarize feature, we perform k-means based clustering and
store the cluster centroids as its fingerprint. The number of
TQ
Query frames
Qoring:Query
Features
P
T
Q
p
Signature:Q
P
M
Query Symbol
sQ1
sQM
TN
Videos frames
Feature Set per
Video
P
T
N
p
Video
Fingerprint
P
FN
VQ Model
Symbol
sX1
sXF1
VQ based Pruned
Search and Return Top
K Neighbors
Naive Linear
Search on Top K
Neighbors using
Qorig
Return best match
Vi*, Decide
whether the best
match is a
duplicate
Database video Query Video

Volume: 5 Issue: 11 139 – 142
_______________________________________________________________________________________________
141
_______________________________________________________________________________________
0
20
40
60
80
100
Video1 Video 2 Video 3 Video 4 Video 5
CLD
OCS
clusters is fixed at a certain fraction of video frames.
Therefore, the fingerprint size varies with the video length. K-
means based clustering generally produces compact video
signatures, perform VQ based encoding on the signatures to
obtain sparse histogram-based representations, compute the
relevant keyframes, and then perform two-stage search to
return the best matched model video.
2) OCS:Perception of color is usually not best represented
in RGB. A better model of HVS is the so-call opponent color
model. In [14], the opponent histogram is a combination of
three 1-D histograms based on the channels of the opponent
color space. The intensity is represented in channel O3 and the
color information is in channels O1 and O2. Due to the
subtraction in O1 and O2, the offsets will cancel out if they are
equal for all channels (e.g. a white light source). Therefore,
these color models are shift-invariant with respect to light
intensity. The intensity channel O3 has no invariance
properties. The histogram intervals for the opponent color
space have ranges different from the RGB model.
C. Distance Mesure
The duplicate detection task is to retrieve the best matching
model video fingerprint for a given query fingerprint. The
model-to-query distance is computed using a non-metric
distance measure between the fingerprints. The distance
function to compare a (Fi × p) sized model fingerprint Xi with
the (M × p) sized query signature Q is d(Xi ,Q).
d(Xi, Q) = {M
k=1 ║Xj
i
- Qk
Where ║Xj
i
− Qk║ refers to the L1 distance between Xj
i
,
the jth
feature vector of Xi
and Qk , the kth
feature vector of Q.
Previous technique effectively depend on a single query frame
and model video frame, and errors occur when this query (or
model) frame is not representative of the query (or model)
video. In distance function (1), d (Xi
,Q) is computed
considering all the ―minimum query frame-to-model video‖
terms and hence, the effect of one (or more) mismatched query
feature vector is compensated. The summation of the best-
matching distance of each vector in Q with all the vectors in
the signature for the original video (Xi
) will yield a small
distance. Hence, the model-to-query distance is small when
the query is a (noisy) subset of the original model video.
D. Video Matching
To perform efficient search, we propose a two phase
procedure. The first phase is a coarse search to return the top-
K nearest neighbors (NN) from all the N model videos. The
second Phase uses the unquantized features for the top-K NN
videos to find the best matching video Vi*.The NLS algorithm
implements the two-pass method without any pruning. In the
first pass, it retrieves the top-K candidates based on the
smaller query signature Q by performing a full dataset scan,
the first pass compares the model fingerprint Xi
with the query
signature Q, and returns the K nearest videos. The second pass
finds the best matched video Vi from these K videos, based on
the larger query signature Qorig.
When the feature vectors are vector quantized, an inter-
vector distance reduces to an inter-symbol distance, which is
fixed once the VQ codevectors are fixed. Hence, we vector
quantize the feature vectors and represent the signatures as
histograms, whose bins are the VQ symbol indices. For a
given VQ, we store and pre-compute the inter-symbol distance
matrix in memory. Describe the VQ-based signature creation
[1]. Using the CLD features extracted from the database video
frames, a VQ of size U is constructed using the Linde Buzo
Gray (LBG) algorithm [10].
1) Algorithm
Step1: Query video is converted to number of frames.
Step2: Feature is extracted from frames of query videos by
using CLD method.
Step3: The model-to-query distance is computed using
distance measure
Step4: Coarse search return top k-NN from all the model
videos
Step5: refined search find the best matching video Vi*
Step6: decides whether the query is indeed a duplicate
derived from Vi*.
Duplicate confirmation, After finding the best matched
video Vi*, to confirm whether it is indeed a duplicate use
threshold approach on the model-to-query distance.
IV. EXPERIMENTAL RESULT
We describe the duplicate detection experiment for feature
comparison. A database contains 100 videos, worth about 1
hours of video content. 20 original videos are used to generate
the query videos. We use various image processing and noise
addition operations to generate the queries. gamma correction,
JPEG compression at quality different factors, varying frame
rates, We have experimented with different query lengths, as a
duplicate can be constructed as a subset of a model video full-
length, 35% of the original model video length.
We also perform experiment on video queries collected
form YouTube. These web videos are identical or
approximately identical videos close to the exact duplicate of
each other, but different in file formats, encoding parameters,
photometric variations (color, lighting changes), editing
operations (caption, logo and border insertion), different
lengths, and certain modifications (frames add/remove).
TABLE I. COMPARISON OF THE SIMILARITY OF CLD AND OCS
Similarity value
Input videos CLD OCS
Video1 65.21 76.19
Video 2 79.5 87.14
Video 3 69.28 74.78
Video 4 71.0 81.0
Video 5 64.70 87.5
Figure 2. Similarity between videos with CLD and OCS

Volume: 5 Issue: 11 139 – 142
_______________________________________________________________________________________________
142
_______________________________________________________________________________________
V. CONCLUSION
This paper can deal with various kinds of video
transformations, such as video compression, blurred, video
cutting. As well, two feature extraction methods are used for
extract feature from video, and find the similarity between
videos using distance measure, and indexing method used to
speed up the matching process. We retrieve the duplicate
video, an average detection accuracy of over 97% when the
query video as a noisy subset of the original video, and 80%
detection accuracy when the query videos are real time videos.
REFRENCES
[1] Anindya Sarkar, Vishwarkarma Singh, Pratim Ghosh, B. S.
Manjunath, and Ambuj Singh. Efficient and Robust Detection of
Duplicate Videos in a Large Database,2011.
[2] X. Wu, A. G. Hauptmann, and C. Ngo. Practical elimination of
near-duplicates from web video search. In Proceedings of the
15th
International Conference on Multimedia, pages 218–227.
ACM, 2007.
[3] S. Cheung and A. Zakhor. Estimation of web video multiplicity.
In Proc. SPIE–Internet Imaging, volume 3964, pages 34–36,
1999.
[4] L. Liu, W. Lai, X. Hua, and S. Yang. Video Histogram: A Novel
Video Signature for Efficient Web Video Duplicate Detection.
Lecture Notes in Computer Science, 4352:94–103, 2007.
[5] J. Law-To, L. Chen, A. Joly, I. Laptev, O. Buisson, V. Gouet-
Brunet, N. Boujemaa, and F. Stentiford. Video copy detection: a
comparative study. In Proc. of CIVR, pages 371–378. ACM,
2007.
[6] E. Kasutani and A. Yamada. The MPEG-7 color layout
descriptor: a compact image feature description for high-speed
image/video segment retrieval. In Proc. of ICIP, volume 1, pages
674–677, 2001.
[7] M. Bertini, A. D. Bimbo, and W. Nunziati. Video clip matching
using MPEG-7 descriptors and edit distance. In Proc. of CIVR,
pages 133–142, 2006.
[8] A. Joly, C. Frelicot, and O. Buisson. Robust content-based
video copy identification in a large reference database. In Int.
Conf. onImage and Video Retrieval, pages 414–424, 2003.
[9] H. Lejsek, F. H. Asmundsson, B. Jonsson, and L. Amsaleg. NV-
tree: An efficient disk-based index for approximate search in
very large high-dimensional collections. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 99(1), 2008.
[10] K. Dadason, H. Lejsek, F. Asmundsson, B. Jonsson, and L.
Amsaleg. Videntifier: identifying pirated videos in real-time. In
Proc. Of the 15th International Conference on Multimedia, pages
471–472. ACM, 2007.
[11] A. Joly, O. Buisson, and C. Frelicot. Statistical similarity search
applied to content-based video copy detection. Int. Conf. on
Data Engineering Workshops, page 1285, 2005.
[12] V. Athitsos, M. Potamias, P. Papapetrou, and G. Kollios.
Nearest neighbor retrieval using distance-based hashing. Proc. of
ICDE, pages 327–336, April 2008.
[13] A. Joly, O. Buisson, and C. Frelicot. Content-based copy
retrieval using distortion-based probabilistic similarity search.
Multimedia,IEEE Transactions on, 9(2):293–306, Feb. 2007.
[14] Koen E. A. van de Sande and Theo Gevers and Cees G. M.
Snoek. Evaluation of Color Descriptors for Object and Scene
Recognition. IEEE Transactions on Pattern Analysis and
Machine Intelligence, volume 32, pages1582 -1596.

Efficient and Robust Detection of Duplicate Videos in a Database

More Related Content

What's hot (17)

Similar to Efficient and Robust Detection of Duplicate Videos in a Database (20)

More from rahulmonikasharma (20)

Recently uploaded (20)

Efficient and Robust Detection of Duplicate Videos in a Database