An Improved Feature Vector for Content-based Image Retrieval in DCT
Domain
Cong Bai, Kidiyo Kpalma and Joseph Ronsin
Universit´e Europ´enne de Bretagne, INSA de Rennes, IETR, UMR 6164, Rennes, France
Keywords:
Content-based Image Retrieval, DCT, Texture, Face Recognition.
Abstract:
This paper proposes an improved approach for content-based image retrieval in Discrete Cosine Transform
domain. For each 4x4 DCT block, we calculate the statistical information of three groups of AC coefficients
and propose to use these values to form the AC-Pattern and use DC coefficients of neighboring blocks to con-
struct DC-Pattern. The histograms of these two patterns are constructed and their selections are concatenated
as feature descriptor. Similarity between the feature descriptors is measured by χ
2
distance. Experiments
executed on widely used face and texture databases show that better performance can be observed with the
proposal compared with other classical method and state-of-the-art approaches.
1 INTRODUCTION
As a transform adopted in JPEG compression stan-
dard, Discrete Cosine Transform (DCT) is a powerful
tool to extract features from images. Consequently
in last decades, many researches appeared in image
retrieval based on DCT. DCT has the capability to
compact the energy, i.e. much of the energy lies in
low frequency coefficients, so high frequency can be
discarded without visible distortion. In other words, a
reduced part of DCT coefficients can efficiently rep-
resent the image contents. In comparison to the use
of all of the coefficients, this consideration reduces
the complexity and redundancy of the feature vectors
that are generated for image retrieval. Furthermore,
the AC coefficients of some regions represent some
directional information.
Different compositions of the coefficients for a
feature vector construction have been proposed. In
(Tsai et al., 2006), the upper left coefficients of each
8x8 block are categorized into four groups: one is
DC coefficient and other three includes the AC coef-
ficients which have vertical, horizontal and diagonal
information. These four groups compose the feature
vectors. In (Zhong and Def´ee, 2005), 16 coefficients
from each 4x4 block are used to construct 2 patterns
and then their histograms are constructed as descrip-
tors to do face recognition. In (Bai et al., 2012), a
selection of coefficients is done and the feature de-
scriptors are constructed from the histogram of this
selection. Our approach proposed in this paper is in-
spired from these previous works and those consider-
ations mentioned in last paragraph.
In this paper, we present a simple but effective
way to construct the feature vectors, and use the part
of histogram of these vectors as the descriptors of
the images. Similarity of the descriptors between
query and images in the database is measured by χ
2
distance. Experimental results show that the pro-
posed method can apply both on face database and
texture database and can achieve better performance
than the referred methods including state-of-the-art
approaches.
The rest of the paper is organized as follows: the
principle of constructing feature vectors is presented
in section 2. Section 3 gives the process of forming
feature descriptors and selection of similarity mea-
surement. Experimental results are shown in section
4 and conclusion is given in section 5.
2 FEATURE VECTOR
Images are transformed by DCT firstly. In this study,
we use 50% overlapping 4x4 DCT block transform.
To eliminate the effect of luminance variations, lu-
minance normalization needs to be done before con-
structing AC and DC patterns. We adopt the method
presented in (Zhong and Def´ee, 2005) as a pre-
processing step. From these pre-processed coeffi-
cients, AC-Patterns and DC-Patterns are constructed.
742
Bai C., Kpalma K. and Ronsin J..
An Improved Feature Vector for Content-based Image Retrieval in DCT Domain.
DOI: 10.5220/0004206607420745
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 742-745
ISBN: 978-989-8565-47-1
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
For each block, AC-Pattern is formed by 9 coeffi-
cients and DC-Pattern is constructed by the DC co-
efficients of the block itself and those of its 8 neigh-
boring blocks. So intra-block features and inter-block
features can be represented by these two patterns.
2.1 AC-Pattern
In DCT domain, energy is packed into a few coeffi-
cients, and some coefficients represent directional in-
formation. So the proposed approach selects 9 co-
efficients out of all 15 AC coefficients in each block
and uses their statistical information to construct the
AC-Pattern. These 9 coefficients are categorized into
3 groups: horizontal (Group H), vertical (Group V)
and diagonal (Group D). For each group, the sum
of the coefficients is calculated firstly and then the
squared-differences between each coefficient and the
sum of this group are calculated. Finally, the sums
of squared-differences of each group are used to con-
struct AC-Pattern. The process of forming AC-Pattern
is shown in Figure 1. Compared with the method
of (Zhong and Def´ee, 2005), this way of construct-
ing AC-Pattern reduces obviously the complexities of
the feature vector. For example, the maximal dimen-
sion of the AC-Pattern of (Zhong and Def´ee, 2005) is
15 and the one of ours is only 3. This number remains
the same as the method of (Bai et al., 2012).
¦
=
−=
3,2,1
2
11
)(
i
i
SC
δ
¦
=
−=
12,8,4
2
22
)(
i
i
SC
δ
¦
=
−=
15,10,5
2
33
)(
i
i
SC
δ
1
δ
2
δ
3
δ
Figure 1: AC-Pattern construction: (a) Three groups of
AC coefficients are extracted from DCT block (b) Sum of
group (c) Sum of squared-differences (d) AC-Pattern.
2.2 DC-Pattern
Different from previous AC-Patterns that describe lo-
cal information inside each block, DC-Pattern will
consider global information by using gradients be-
tween DC coefficients of each block and those of
its neighbors blocks. So DC-DirecVec (Zhong and
Def´ee, 2005) is used as DC-Patterns.
3 FEATURE DESCRIPTOR AND
SIMILARITY MEASUREMENT
3.1 Feature descriptor
In this study, we use the histogram of patterns instead
of histogram of individual DCT coefficients as feature
descriptor. The histogram of AC-Patterns and DC-
Patterns are defined as the number of appearance of
patterns in DCT domain. A disadvantage of the his-
togram method is that it requires a large number of
histogram bins, typically several thousands, to cap-
ture accurately information of feature vector. Thus it
leads to complexity in both storage of image indices
and retrieval timing. To overcome this drawback, we
adopt two modifications.
From the original histogram of AC-Patterns we
can make two observations and then two improve-
ments are respectivelyadopted: the first is that there is
only a few part of AC-Patterns which appears in large
quantities and a large number of AC-Patterns that ap-
pears rarely (Zhong and Def´ee, 2005). So in consider-
ation of time-consuming and efficiency, we just select
ACbins AC-Patterns which have highest occurrence
to construct the histogram. The second observation is
that the first AC-Pattern inside the histogram is very
dominant. This AC-Pattern corresponds to uniform
blocks in the image and consequently we will discard
it. Thus we obtain the AC-Pattern histogram H
AC
, as
show in Figure 2. In this histogram, ACbins = 70.
0 10 20 30 40 50 60 70
0
20
40
60
80
100
120
140
160
180
200
AC-Pattern
Number of AC-Pattern
Histogram of AC-Pattern after selection
Figure 2: Histogram of AC-Patterns after selection.
For DC-Pattern histogram, we can also make the
same observation that only a part of DC-Patterns
occurs in large quantities while other part of
DC-Patterns appears rarely. So we select DCbins
dominant DC-Patterns to construct DC-Pattern his-
togram H
DC
.
Finally, we use the concatenation of AC-Pattern
and DC-Pattern histograms to build the descriptors.
In this context, the descriptor is defined as follows:
D = [(1− α)× H
AC
, α× H
DC
] (1)
AnImprovedFeatureVectorforContent-basedImageRetrievalinDCTDomain
743
where α is a weight parameter that controls the impact
of AC-Patterns and DC-Patterns histogram.
3.2 Similarity Measurement
The similarity between query and images in the
database is assessed by the distance between feature
descriptors. A similarity measure assigns a lower dis-
tance or high score to more similar objects.
In our approach, χ
2
distance is used to measure
the similarity. Assuming D
i
and D
j
are the query and
target feature descriptors respectively, the χ
2
distance
is defined as follows:
Dis
i,j
=
m
∑
k=1
(D
i
(k) − D
j
(k))
2
D
i
(k) + D
j
(k)
(2)
where m indicates the dimension of descriptors.
4 EXPERIMENTAL RESULTS
To evaluate the proposal, we perform experiments on
ORL face database (AT&T Laboratories Cambridge,
1992), GTF face database (Georgia Tech, 1999),
FERET face database (Phillips et al., 2000) and Vis-
Tex texture database (Media Laboratory, 1995).
4.1 Experiments on Face Databases
The ORL database includes 10 different images of
40 persons. The GTF database includes 15 differ-
ent face images of 50 different persons. As we want
to compare our proposal with the methods presented
in (Zhong and Def´ee, 2005) and (Bai et al., 2012), we
implemented similar experiments as them, that means
use first 6 images as image database and remaining
4 images as query images for recognition on ORL
database and first 11 images of each person as image
database and remaining 4 images as query images on
GTF database.
As ORL and GTF are relatively small databases,
to further illustrate the contribution of the proposal,
we also implement it on FERET database which con-
tains more than 10000 images from more than 1000
persons. The two sets of frontal view faces fa and fb
were selected to evaluate the proposed method: fb is
used as query images for retrieval from the fa.
For evaluating the performance, we use Equal Er-
ror Rate (EER) (Bolle et al., 2000). Images are con-
sidered as similar if the distance between their fea-
tures descriptors is under a given threshold. So con-
sidering a query image belonging to class A, two
things could occur: on one hand, it could be recog-
nized rightly; on the other hand, it could be falsely
rejected from class A, then the ratio of how many im-
ages of class A are in this situation is called False Re-
jected Rate (FRR). In contrast, considering a query
image out of class A, when it is compared with the
images of class A, it could be rejected rightly or it
could be falsely accepted as class A, then the ratio
of how many images of other classes are in this situa-
tion is defined as False Accept Rate (FAR). These two
rates will change when the threshold changes. When
FRR and FAR take equal values, an equal error rate
(EER) is got. The lower the EER is, the better is the
performance.
As mentioned before, the concatenation of the
AC-Pattern histogram and the DC-Pattern histogram
is used to do image retrieval. To verify the effec-
tiveness of the proposed new method on constructing
AC-Pattern, the DC-Patterns histogram is the same
through all the experiments, but the construction of
AC-Pattern are different. We name the method pre-
sented in (Zhong and Def´ee, 2005) as “Zhong”, the
method presented in (Bai et al., 2012) as “Bai”. For
these two methods and our proposal, we tested differ-
ent sets of parameters to find the one that can assure
the best performance, and only the best performances
are compared that are shown in Table 1. From this
table, we can see that the proposal outperforms the
referred methods.
Table 1: Comparison of EER (%) on face databases.
Database/Method Zhong Bai Proposal
ORL 6.07 3.75 3.25
GTF 11.57 11.19 9.84
FERET 3.75 3.57 3.35
4.2 Experiments on Texture Database
In order to evaluate extensibility of the proposal in a
wider application field, we work on a selection of im-
ages from the Vistex Texture Database (Media Lab-
oratory, 1995), consisting of 40 textures which have
already been extensively used in texture image re-
trieval literature (Do and Vetterli, 2002) (Kokare M.
and Chatterji, 2005) (Kwitt and Uhl, 2010).
The 512 × 512 pixels color version of the tex-
tures are divided into 16 non-overlapping subimages
(128×128 pixels) and convertedto gray scale images,
thus creating a database of 640 images belonging to
40 texture-classes, each class includes 16 different
samples. In this retrieval experiments, each image is
used as query image. The relevant images for each
query consists of all the subimages from the same
original texture. Like in other literatures (Do and Vet-
terli, 2002) (Kokare M. and Chatterji, 2005) (Kwitt
and Uhl, 2010), we use the average retrieval rate
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
744
(ARR) to evaluate the performance. For a given query
image, and a given set of retrieved images, the re-
trieval rate is defined as the percentage of the num-
ber of correct images retrieved on the total number
of retrieved images. For comparison purpose, top-16
match images are retrieved for each query.
Table 2: Comparison of ARR on Vistex.
Method RCWF CWT PTR CWT+RCWF Proposal
ARR(%) 75.78 80.78 81.73 82.34 84.04
Table 2 provides an objective comparison of
ARR. In this table, RCWF indicates Rotated Com-
plex Wavelet Filters method proposed in (Kokare M.
and Chatterji, 2005). CWT represents the Com-
plex Wavelet Transform method presented in (Kings-
bury, 1999). CWT+RCWF is a mixed method
also presented in (Kokare M. and Chatterji, 2005).
PTR (Kwitt and Uhl, 2010) is a probabilistic texture
retrieval method based on dual-tree complex wavelet
transform. From this table, it can be observed that our
proposal outperforms other methods.
5 CONCLUSIONS
We have presented an improvedapproach for content-
based image retrieval in DCT domain. A new way to
construct AC-Patterns is proposed: this proposal uses
the statistical information of every directional groups
of AC coefficients. Compared with other methods of
constructing AC-Patterns (Zhong and Def´ee, 2005),
the proposed method reduces the dimension of AC-
Patterns. And we also evaluated the extensibility of
our proposal by applying it on two different kinds
of database: face database, which has structural con-
tents, and texture database, which has both structural
and unstructured contents. Our approach is evalu-
ated on widely used ORL, GTF, FERET and VisTex
database. The experimental results show that our ap-
proach outperforms the referred methods, including
state-of-the-art methods.
REFERENCES
AT&T Laboratories Cambridge (1992). ORL database.
http:// www.cl.cam.ac.uk/research/dtg/attarchive/face
database.html. Online; accessed March 2010.
Bai, C., Kpalma, K., and Ronsin, J. (2012). A new descripor
based on 2D DCT for image retrieval. In Proceedings
of the International Conference on Computer Vision
Theory and Applications, pages 714–717.
Bolle, R., Pankanti, S., and Ratha, N. (2000). Evaluation
techniques for biometrics-based authentication sys-
tems (FRR). In Pattern Recognition, 2000. Proceed-
ings. 15th International Conference on, volume 2,
pages 831 –837 vol.2.
Do, M. and Vetterli, M. (2002). Wavelet-based tex-
ture retrieval using generalized gaussian density and
kullback-leibler distance. Image Processing, IEEE
Transactions on, 11(2):146 –158.
Georgia Tech (1999). GTF database. http://
www.anefian.com/research/face reco.htm. Online;
accessed March 2010.
Kingsbury, N. G. (1999). Image processing with complex
wavelet. Phil. Trans. Roy. Soc., 357:2543–2560.
Kokare M., B. P. and Chatterji, B. (2005). Texture image
retrieval using new rotated complex wavelet filters.
IEEE Transactions on Systems, Man, and Cybernet-
ics, Part B: Cybernetics, 35(6):1168–1178.
Kwitt, R. and Uhl, A. (2010). Lightweight probabilistic tex-
ture retrieval. Image Processing, IEEE Transactions
on, 19(1):241 –253.
Media Laboratory, M. (1995). Vistex database of tex-
tures. http:// vismod.media.mit.edu/vismod/imagery/
VisionTexture/. Online; accessed Dec. 2010.
Phillips, P., Moon, H., Rizvi, S., and Rauss, P. (2000). The
FERET evaluation methodology for face-recognition
algorithms. Pattern Analysis and Machine Intelli-
gence, IEEE Transactions on, 22(10):1090 – 1104.
Tsai, T., Huang, Y.-P., and Chiang, T.-W. (2006). Image re-
trieval based on dominant texture features. In Indus-
trial Electronics, 2006 IEEE International Symposium
on, volume 1, pages 441 –446.
Zhong, D. and Def´ee, I. (2005). DCT histogram optimiza-
tion for image database retrieval. Pattern Recognition
Letters, 26(14):2272 – 2281.
AnImprovedFeatureVectorforContent-basedImageRetrievalinDCTDomain
745
