Quantifying the Specificity of Near-duplicate
Image Classification Functions
Richard Connor
and Franco Alberto Cardillo
Department of Computer and Information Sciences, University of Strathclyde, Glasgow, G1 1XH, Scotland
Istituto di Scienza e Tecnologie dell’Informazione, Consiglio Nazionale delle Ricerche, Pisa, Italy
Near-duplicate Image Detection, Benchmark, Image Similarity Function, Specificity, Forensic Image Detec-
There are many published methods for detecting similar and near-duplicate images. Here, we consider their
use in the context of unsupervised near-duplicate detection, where the task is to find a (relatively small) near-
duplicate intersection of two large candidate sets. Such scenarios are of particular importance in forensic
near-duplicate detection. The essential properties of a such a function are: performance, sensitivity, and
specificity. We show that, as collection sizes increase, then specificity becomes the most important of these,
as without very high specificity huge numbers of false positive matches will be identified. This makes even
very fast, highly sensitive methods completely useless. Until now, to our knowledge, no attempt has been
made to measure the specificity of near-duplicate finders, or even to compare them with each other. Recently,
a benchmark set of near-duplicate images has been established which allows such assessment by giving a
near-duplicate ground truth over a large general image collection. Using this we establish a methodology for
calculating specificity. A number of the most likely candidate functions are compared with each other and
accurate measurement of sensitivity vs. specificity are given. We believe these are the first such figures be to
calculated for any such function.
In forensic image detection, it is commonly required
to determine if one large image collection contains
images for which near-duplicates exist within another
large collection. For example, the UK National Crime
Agency has a collection of approximately 10
child abuse images. A media device seized from a
suspect may contain 10
images, but perhaps only 10
of these would be of child abuse and thus a subject for
prosecution. A knowledgeable suspect will have per-
formed minor visual editing of these images, making
them detectable only by near-duplicate finding func-
Finding images similar to one other, from within
a huge collection, is nowadays a (relatively) solved
problem, and has been shown to scale to up to col-
lections of well over 10
images. However almost
all such results are set in a human-guided search sce-
nario, where a single image is presented to the system
and a human is available to pick the best results from
those returned. Here we examine quite a different sce-
nario, where two moderately large collections are the
input, and the desired output is the set of images of
which near-duplicates exist across the two collections.
This causes new problems, due to the inherent squar-
ing: if both collections contain a relatively modest 10
images, then there are 10
pairs to consider, rather
more comparisons than required to compare a single
image against Google’s indexed image collection.
The performance issues can be handled by various
means, including using the obvious parallelisation in-
herent in the problem. In this article we concentrate
on a more subtle problem, that of the precision, or
specificity, of the classification function. As there are
a really huge number of pairs of images to consider,
the specificity must be incredibly high to avoid very
large numbers of false positives. For example, speci-
ficity of 1
would, in most contexts, give excel-
lent precision; but set in this example context would
give 10
false positives, impossible for any human to
check. This is greatly in contrast with human-guided
search, where specificity as low as 0.5 is quite accept-
Finding near-duplicate images in this context re-
quires a classification function; that is, a boolean sim-
Connor, R. and Cardillo, F.
Quantifying the Specificity of Near-duplicate Image Classification Functions.
DOI: 10.5220/0005785406470654
In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 647-654
ISBN: 978-989-758-175-5
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
ilarity function s which allows the detection of the
near-duplicate intersection: from collections X and Y ,
the near-duplicate intersection is defined as the set of
pairs {x, y X,Y , s(x, y)}. As similarity functions
are generally numeric, a decision must be made for
a threshold at which to apply the function, giving ac-
ceptable tradeoffs for sensitivity and specificity.
Specificity however is very difficult to measure,
requiring as it does a very large collection with a
known ground truth. Such a collection has been es-
tablished for images in one context, in (Connor et al.,
2015; Connor, 2015), which gives a quantified esti-
mate of the near-duplicate ground truth for a collec-
tion of one million images. Using this collection, we
show a methodology for quantifying the specificity of
near-duplicate finders, and give results for some of the
most likely candidate functions.
Finding a pair of images, one of which has been cre-
ated by applying minor transformations to the other,
is not an easy problem. Such transformations, avail-
able in off-the-shelf image processing software, in-
clude changes to contrast, brightness, colour, texture,
and sharpness. It is easy for the human brain to see
that two such images are obviously the same modulo
such changes, but these are exactly the properties that
most similarity functions rely on for a mechanised
quantification of similarity.
Implicitly, the context of testing for similarity
function performance is usually a human user search-
ing a very large collection for images that are most
similar to a given reference image. As such, efforts
tend to be concentrated on the performance and recall
of nearest-neighbour search, and success is generally
measured in terms of how many correct results can be
obtained from a huge collection within a short time.
Our context of interest is quite different however.
Increasing pressure is being put on Internet Service
Providers, social network and search engine providers
to filter image and video content that is being used il-
legally, for example the portrayal of child abuse or
content which is subject to copyright. Such content is
always subject to minor changes, for a number of rea-
sons but increasingly including deliberate attempts to
evade detection. Therefore near-duplicate rather than
duplicate detection is increasingly required.
In these and other contexts, the requirement is to
automatically search large numbers of images against
a reference collection which is also large, for exam-
ple 10
to 10
images or keyframes. When a pos-
sible near-duplicate is detected, the two images then
Figure 1: Very similar images, as determined by edge his-
Figure 2: Very similar images, as determined by colour his-
require to be tested by human inspection. Efficiency
and high-quality semantic matching are requirements,
but false positive detection becomes a much bigger is-
sue. As the number of comparisons required is huge,
and the majority of candidates have no match, false
positive detection must be an incredibly rare event.
Figures 1 and 2 show some perhaps surprising
motivating examples where very small distances have
been observed over very different images, as a result
of a very low false positive probability being applied
over a very large sample population. In both of these
cases, the image pairs shown have very small dis-
tances measured according to a particular similarity
function: that is, distances that were smaller than the
majority of true near-duplicate pairs. These examples
were found in a collection of 10
image pairs: such
coincidental matches are quite likely to occur even if
their probability is as low as 10
The contribution of this paper is to show a way
of performing specificity measurement useful for this
context, and to give some early analysis of various
near-duplicate finding mechanisms.
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
We are not aware of any other work which attempts
to measure comparative sensitivity and specificity for
different image similarity functions. We believe one
reason for this is the lack of any large collection of im-
ages with a known ground truth of similarity, a lack
mentioned by various authors, for example: “We do
not have access to ground-truth data for our exper-
iments, since we are not aware of any large public
corpus in which near duplicate images have been an-
notated. (Chum et al., 2007), and the same lack is
noted in (Jinda-Apiraksa et al., 2013). (Vonikakis
et al., 2014) note “ Although the target application of
this dataset is image retrieval, it was selected due to
the lack of other appropriate datasets [. ..]”.
There is also only a little work which objectively
compares different image similarity functions. Royo
(Ventura Royo, 2010) gives a comparison of differ-
ent MPEG-7 techniques for image search, however,
in common with most published work in this domain,
the notion of success is based on correct retrieval from
a relatively small collection of known images. Inter-
estingly, the author finds that MPEG-7 Colour Struc-
ture gives the best performance in this context, while
we find it by far the worst of those tested. The main
point is that the issues addressed here are defined by
the scale of the search, and comparative studies over
small collections of images do not give useful results.
Douze et al. (Won et al., 2002) compare the
GIST (Oliva and Torralba, 2001) image characterisa-
tion with a “bag-of-features” approach, and find it su-
perior for near-duplicate images among other things.
They test over significant sizes of image collection
by adding their ground truth of ‘similar’ images to
large ‘distractor’ sets, including MIR-Flickr. How-
ever they generate near-duplicate images from a rela-
tively small, chosen set, through programmed quality
loss and cropping, and also apply strong transforma-
tions which are not covered by our definition of near-
duplicate. The use of generated near-duplicate images
weakens the results, as for any known transformation
it is relatively straightforward to predict which simi-
larity functions will perform well.
Foo et al. (Foo et al., 2006) gave one of the ear-
liest treatments of the issue of near-duplicate image
finding as a subject in its own right, and defined two
categories of near-duplicate images: IND and NIND.
IND images are “derived from the same digital source
after applying some transformations”, and NIND im-
ages “share the same scenes and objects”. Here we
concentrate primarily on the IND category which is
most appropriate for our defined problem domain.
Our measurement of image similarity functions
for this purpose is based over the MIR-Flickr collec-
tion of one million images (Huiskes and Lew, 2008;
Huiskes et al., 2010). This dataset consists of one mil-
lion “interesting” images downloaded from the web-
site flickr.com through its public API. The “interest-
ingness” of the images represents a score attributed
by the Flickr service by taking into account the com-
ments and the clickthroughs on the images. Since
the 1M images included in the dataset were not se-
lected with a specific task or set of criteria in mind,
they should represent a good benchmark for evalua-
tion of near duplicate detection algorithms on large
image datasets.
Using this collection, the Mir-Flickr Near Dupli-
cate (MFND) classification (Connor et al., 2015; Con-
nor, 2015) identifies three sets of clusters which occur
within the original set of one million images: dupli-
cate clusters, IND near-duplicate clusters, and NIND
near-duplicate clusters. The IND collection comprises
1,958 clusters containing a total of 4,071 images, the
majority of clusters containing only two elements.
Using population statistics, the authors have shown
that these are almost all of the IND pairs that exist
within the whole set. Based on three relatively inde-
pendent near-duplicate finding functions, an estimate
is made of the total number which exist, including
those not found. Furthermore, the standard error of
this estimate has also been established at under 0.02.
Therefore if an image is chosen randomly from out-
side these three clusters, there is a very low probabil-
ity that a near-duplicate will exist within the rest of
the collection. We have used this property in order
to establish the specificity measurements given in this
There are many image similarity functions de-
scribed in the literature. In general, these can be clas-
sified into those relying on global, and local, features.
Here we report only the class of global feature func-
tions, as we believe that these are most likely to be
suitable for our specific problem domain. This is a
questionable assumption that we are investigating fur-
ther, although backed up by results in (Won et al.,
2002) and (Chum et al., 2007). Table 1 summarises
the different characterisations studied.
In many cases a particular distance metric is spec-
ified as a part of the published mechanism. For exam-
ple (Won et al., 2002) prescribes the use of L
hattan) distance, and most assume that L
distance is the best measure of distance between char-
acterisations. However we have found these presump-
tions are most often wrong, certainly in this context,
and we have tested all characterisations over a number
of different metrics.
In some cases, optimisation mechanisms are bun-
Quantifying the Specificity of Near-duplicate Image Classification Functions
Table 1: Image characterisations used.
Abbreviation Characterisation Reference
Cs MPEG-7 Colour Structures (Bober, 2001)
Csl MPEG-7 Colour Structure Layout (ISO-15938, )
Eh MPEG-7 Edge Histograms (Won et al., 2002)
Ghch Global Hierarchical Colour Histograms (Chum et al., 2007)
Gist GIST (Oliva and Torralba, 2001)
Ht MPEG-7 Heterogeneous Textures (Bober, 2001)
Phash Perceptual Hashing (Niu and Jiao, 2008)
dled with the extraction description, for example (Niu
and Jiao, 2008) extracts a bitmap for comparison
with Hamming distance, and (Chum et al., 2007)
uses locality-sensitive hashing techniques over the ex-
tracted representations. Rather than do this we main-
tain the original representations and apply general
metrics over them; optimisation techniques are best
treated separately.
Finally, three of these characterisations rely upon
the image colour palette; Cs very strongly, and Csl
and Ghch much less so, as despite their names the
most significant information in these is extracted ac-
cording to pixel intensity rather than colour. How-
ever, results for Cs are very bad, and based on this
observation we repeated tests on Csl and Ghch using
only intensity, rather than colours. As they were in all
cases better, we report only these variants here. From
observation, many IND near-duplicate images in our
reference set have been produced by making changes
to the colour palette, which explains this observation.
We assume that any near-duplicate finder is based on
a positive numeric function D over any two images.
Normally D will be a proper distance metric to allow
scaling of the search, but this is not an essential se-
mantic property.
To run an unsupervised search, it is necessary to
use D as a classification function. To achieve this,
a distance threshold t must be chosen to be used in
conjunction with D to form a predicate function D
over image pairs, such that D
(x,y) = D(x,y) t.
The problem domain can then be characterised as
the requirement to find the near-duplicate intersection
of two image sets X and Y , based on a concepual
near-duplicate relation ND, where this is defined as
the set of pairs X
Y where (x, y) X
Normal definitions of sensitivity and specificity
for a threshold function D
can now be defined as con-
ditional probabilities:
= P (D
(x,y)),given (x,y) X
= P ((x,y) / X
Y ), given ¬D
In general, as the search threshold t is increased,
the sensitivity increases and the specificity decreases.
As the collections become larger, the specificity
becomes increasingly important. After execution of
the unsupervised process, the number of true positive
matches found will be
· |X
Y |
and the number of false positives will be
(1 spec
) · |X| · |Y |
To put this into a realistic context, a typical situ-
ation for the detection of child abuse images from a
seized hard drive is
|X| = 10
,|Y | = 10
Y | = 10
so apparently excellent figures of, for example,
= 0.999,spec
= 0.999
leads to the detection of almost all the 10
true posi-
tives, but these will be impossible to find among 10
false positives. In fact, in this scenario, the specificity
requires to be as high as 1 10
before only around
half of the detected pairs will be true positives.
To investigate such probabilities requires analy-
sis over large image sets with known ground truths,
where the combination of very rare events and very
large populations can be quantified.
The MFND benchmark set defines three sets of image
clusters, which are subsets of the one million images
in the MIR-Flickr set: one of identical, one of IND
and one of NIND images. Results here use the IND
set as this is more likely to be complete, however we
have run the tests reported here using both IND and
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
NIND sets and the outcomes are quantitatively indis-
To measure the sensitivity of a given metric across
a range of thresholds, a set of pairs comprising the
first two images from each cluster was used, giving a
set of 1,958 pairs of IND images. A histogram of the
distances is constructed, and a cumulative probability
density function constructed from this to give sensi-
tivity for each D
across the range of different t values
for each similarity function D.
To measure specificity of a given metric, a set of
5,000 images was randomly selected, ensuring that
none of these were in any of the IND, NIND, or du-
plicate clusters defined by the benchmark collection.
It is thus a safe assumption that none of these im-
ages have a near-duplicate match within the collec-
tion. For each of these images, its nearest neighbour
from within the 1M collection, along with the distance
to this, was determined. Note that in principle this
requires 5 × 10
distance calculations, and metric in-
dexing techniques were used to achieve tractability.
As nearest neighbours are calculated, the small-
est of the nearest-neighbour distances is therefore the
smallest of approximately 5×10
distances measured
across the two sets. As this distance relates two
images which are not near-duplicates, this gives the
specificity of the function D
, where t is this small-
est distance, as 1 (2 × 10
) for these two sets of
In measuring specificity we make the simplifying
assumption that, at least for the smallest few nearest
neighbours measured from the set, these distances are
the smallest from the whole set of potential distances;
that is, not allowing for a single image to have a sec-
ond neighbour which is a smaller absolute distance
than another’s nearest neighbour. This is likely to be
the case, but more importantly it also captures a more
useful figure for use in unsupervised detection sce-
narios, where a threshold-limited nearest neighbour
search will be conducted.
We calculate results in terms of a cumulative prob-
ability density function, therefore a threshold which
admits the five smallest nearest-neighbour distances,
i.e. a cumulative density of 0.001 over the 5,000 test
distances, corresponds to a specificity of 1 10
The cdf value of 0.01 corresponds to a specificity
of 1 10
which as explained earlier is around the
smallest useful threshold for our problem context:
when comparing 10
images against 10
, around 10
false positives would result, this number varying with
the product of the two collection sizes.
To give an example of this analysis, Figure 3
shows simple histograms giving the outcome of using
Manhattan distance over MPEG-7 edge histograms
Figure 3: Eh/Man Histograms.
(Won et al., 2002). The histograms show the distribu-
tion of distances over (1) known IND near-duplicate
image pairs, and (2) 5,000 randomly selected images
and their nearest neighbours from within the set of 1M
images – this latter of course is shifted a long way to
the left of a histogram showing the distribution of dis-
tances of randomly selected image pairs.
As can be seen, there is significant overlap be-
tween the histograms. Figure 4 shows the same data
displayed as the deduced cumulative probability den-
sity functions. From these, it can be seen that using
this function with a threshold of around 2 will give a
specificity of around 1 (10
) and a sensitivity of
around 0.5, but to achieve an order of magnitude im-
provement in specificity would allow a sensitivity of
only around 0.1.
In order to allow comparisons of the different
characterisation and metrics tested, these results are
shown plotted in ROC curves, to show the essen-
tial tradeoff between sensitivity and specificity as the
search threshold is increased. Figure 5 shows the
same data again plotted as ROC graphs, which is the
form we will use from now on to present results of the
different functions tested.
For each different characterisation shown in Table 1,
we have applied the following proper distance met-
rics: Manhattan distance, Euclidean distance, Cosine
and, where possible, Structural Entropic
Distance (SED). There are of course many other po-
tential metrics which could be tested.
SED is the distance metric defined in (Connor
et al., 2011) and refined in (Connor and Moss, 2012)
The angle between the vectors rather than the comple-
ment of its cosine, which is not a proper metric
Quantifying the Specificity of Near-duplicate Image Classification Functions
Figure 4: Eh/Man Cumulative Probability Density (the
lower graph is a magnification at the origin).
for use in general vector spaces. It is defined over
probability distributions, and as such can only be ap-
plied to characterisations all of whose output val-
ues are positive - notably not including those defined
by discrete cosine transform, namely Csl and Phash.
For all characterisations comprising only positive nu-
meric values, they are normalised to sum to 1 for this
purpose. SED gives the same ranking of outcomes
as the better-known Jensen-Shannon distance (Lin,
1991), but is used here as it has better efficiency prop-
erties for querying metric spaces in this context (Con-
nor and Moss, 2012); the qualitative results shown
here would be exactly the same for both metrics.
Finally it is important to stress that we have exam-
ined only the semantic properties of each image sim-
ilarity function, rather than their relative search effi-
ciency which varies very widely, as does their extrac-
tion time. As mentioned earlier, there are many ways
of optimising both of these and an objective compari-
son of performance is not useful in simple terms.
Figures 6 to 11 show results over the same axes for
Figure 5: Eh/Man ROC graphs. The lower graph is the ex-
treme left-hand part of the upper, as required for this con-
Figure 6: Csl ROC graph for three metrics.
six of the seven characterisations tested. In each case
the specificity range is restricted from 1 to 1 10
as previously explained. As can be extrapolated from
these graphs, at this point the sensitivity improvement
given by a decrease in specificity becomes marginal.
The missing characterisation, MPEG-7 Colour
Structures, performs so badly it is not shown, achiev-
ing a sensitivity in this range of less than 0.02 with
any metric. The importance of measuring these over
very large collections is emphasised by the finding by
other authors that this is the best of the MPEG-7 char-
acterisations over small test collections.
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
Figure 7: Eh ROC graph for four metrics.
Figure 8: Ghch ROC graph for four metrics.
Figure 9: Gist ROC graph for four metrics.
The result of these tests is clear from inspect-
ing the graphs: of the 22 methods tested for near-
duplicate classification, the best compromise between
sensitivity and specificity is the use of Structural
Entropic Distance/Jensen-Shannon Distance over the
GIST image representations, which gives a sensitivity
of around 0.8 for a specificity as high as 1 10
Figure 10: Ht ROC graph for four metrics.
Figure 11: Phash ROC graph for three metrics.
Applying this in the context of a scenario with
two collections each of 10
images, with a near-
duplicate intersection of 10
images, then querying at
the appropriate threshold will return 800 of the near-
duplicate image pairs, and 1,000 false positive pairs.
By comparison, for example, using perceptual hash-
ing with Euclidean distance will return 5,000 false
positive pairs for every 500 true pairs.
The main value of this work is to provide a re-usable
methodology by which different near-duplicate find-
ers can be compared for specificity. The particular
value of our method is that it is based upon a large col-
lection of “naturally” selected images which happen
to contain near-duplicates, rather than a constructed
set. The resulting ground truth should not therefore
be biassed towards any particular function.
Validation of the results however is a further chal-
lenge, at least until such time as another large image
Quantifying the Specificity of Near-duplicate Image Classification Functions
set with a near-duplicate ground truth is identified.
We are currently working with the CoPHiR collec-
tion (Bolettieri et al., 2009) (10
images) to establish
whether the figures produced here are consistent.
The variation among the different distance metrics
is a novel observation. Characterisations are normally
used with either L
or L
distance, whereas in the
majority of cases either Cosine or SED/JSD performs
best. These metrics give a closer match according to
the correlation of values within the characterisations,
rather than differences in their absolute magnitude.
However the differences among all the characterisa-
tions do not seem to suggest any general rules about
the best metric to use in different contexts, which re-
quires further investigation.
We would like to thank Richard Martin and Karina
Kubiak-Ossowska of the University of Strathclyde for
help with access to the ARCHIE-WeSt HPC facilities
necessary to achieve some of the analysis.
Franco Alberto Cardillo was supported by the Na-
tional Research Council of Italy (CNR) for a Short-
term Mobility Fellowship (STM), which funded a
stay at the University of Strathclyde in Glasgow (UK)
where part of this work was done.
Richard Connor was supported by a symmet-
ric National Research Council of Italy (CNR) for a
Short-term Mobility Fellowship (STM), no. 33313,
13/05/2015, which funded a stay at the Consiglio
Nazionale delle Ricerche, Pisa, where the work was
further progressed.
Bober, M. (2001). Mpeg-7 visual shape descriptors. IEEE
Transactions on circuits and systems for video tech-
nology, 11(6):716–719.
Bolettieri, P., Esuli, A., Falchi, F., Lucchese, C., Perego,
R., Piccioli, T., and Rabitti, F. (2009). Cophir: a test
collection for content-based image retrieval. CoRR,
Chum, O., Philbin, J., Isard, M., and Zisserman, A. (2007).
Scalable near identical image and shot detection. In
Proceedings of the 6th ACM international conference
on Image and video retrieval, pages 549–556. ACM.
Connor, R. (2015). Mir-flickr near-duplicate data. mir-
Connor, R., Cardillo, F., MacKenzie-Leigh, S., and Moss,
R. (2015). Identification of mir-flickr near-duplicate
images. In 10th International Conference on Com-
puter Vision Theory and Applications.
Connor, R. and Moss, R. (2012). A multivariate correla-
tion distance for vector spaces. In Navarro, G. and
Pestov, V., editors, Similarity Search and Applica-
tions, volume 7404 of Lecture Notes in Computer Sci-
ence, pages 209–225. Springer Berlin Heidelberg.
Connor, R., Simeoni, F., Iakovos, M., and Moss, R. (2011).
A bounded distance metric for comparing tree struc-
ture. Inf. Syst., 36(4):748–764.
Foo, J., Sinha, R., and Zobel, J. (2006). Discovery of image
versions in large collections. In Cham, T.-J., Cai, J.,
Dorai, C., Rajan, D., Chua, T.-S., and Chia, L.-T., edi-
tors, Advances in Multimedia Modeling, volume 4352
of Lecture Notes in Computer Science, pages 433–
442. Springer Berlin Heidelberg.
Huiskes, M. J. and Lew, M. S. (2008). The MIR Flickr
retrieval evaluation. In MIR ’08: Proceedings of the
2008 ACM International Conference on Multimedia
Information Retrieval, New York, NY, USA. ACM.
Huiskes, M. J., Thomee, B., and Lew, M. S. (2010). New
trends and ideas in visual concept detection: The MIR
Flickr retrieval evaluation initiative. In MIR ’10: Pro-
ceedings of the 2010 ACM International Conference
on Multimedia Information Retrieval, pages 527–536,
New York, NY, USA. ACM.
ISO-15938. Mpeg-7 multimedia content description inter-
Jinda-Apiraksa, A., Vonikakis, V., and Winkler, S. (2013).
California-nd: An annotated dataset for near-duplicate
detection in personal photo collections. In Quality of
Multimedia Experience (QoMEX), 2013 Fifth Interna-
tional Workshop on, pages 142–147. IEEE.
Lin, J. (1991). Divergence measures based on the shannon
entropy. Information Theory, IEEE Transactions on,
Niu, X.-m. and Jiao, Y.-h. (2008). An overview of percep-
tual hashing. Acta Electronica Sinica, 36(7):1405–
Oliva, A. and Torralba, A. (2001). Modeling the shape
of the scene: A holistic representation of the spatial
envelope. International Journal of Computer Vision,
Ventura Royo, C. (2010). Image-based query by example
using mpeg-7 visual descriptors.
Vonikakis, V., Jinda-Apiraksa, A., and Winkler, S. (2014).
Photocluster - a multi-clustering technique for near-
duplicate detection in personal photo collections. In
Proc. of the 9th International Conference on Com-
puter Vision Theory and Applications, pages 153–161.
Won, C. S., Park, D. K., and Park, S.-J. (2002). Efficient use
of mpeg-7 edge histogram descriptor. Etri Journal,
VISAPP 2016 - International Conference on Computer Vision Theory and Applications