Improving CBIR Using Feature Extraction Based on
Wavelet Transform
Carolina W. Silva, Pedro H. Bugatti, Marcela X. Ribeiro, Caetano Traina Jr., Agma J. M. Traina
Computer Science Department - ICMC-USP
Caixa Postal 688 - 13560-970 - So Carlos - SP, Brasil
{carolina,pbugatti,mxavier,caetano,agma}@icmc.usp.br
ABSTRACT
The gap semantic and the curse of dimensionality are two short- to support high-level queries.
comings of content-based image retrieval techniques that rely on
automatic feature extracted from images to process similarity queries.
The first one represents the semantic gap that exists between low-
level features automatically extracted by a computational system,
and the high-level user interpretation of images. The second one
involves problems occurring when similarity is defined over high-
dimensional feature spaces. This paper shows a method that deals
with these both shortcomings. We use discrete wavelet transforms
to obtain the image representation from a multiresolution point of
view. The feature vectors were composed of the features from the
approximation subspace, which succinctly represent the images in
the processing of similarity queries. In addition, the multiresolution
method was used to reduce the dimensionality of the feature space.
This work shows the evaluation of three different image datasets,
where the first two are composed of medical images and the third
one is a generic image dataset. The results are promising and show
an improvement of up to 90% for recall values up to 65%, in the
query results using the Daubechies wavelet transform.
Categories and Subject Descriptors
H.3.m [Information Storage and Retrieval]: Miscellaneous; I.4.7
[Image Processing and Computer Vision]: Feature Measurement
feature representation
Keywords
Wavelets, CBIR, Image Processing, Feature Vector, Multiresolu-
tion
1. INTRODUCTION
Content-based image retrieval (CBIR) has been an active research
area in the last fifteen years, and a variety of techniques have been
developed. However, retrieving images based on low-level features
This work has been supported by CNPq, FAPESP and CAPES.
has been proved to be unsatisfactory. New techniques are needed
One of the uses of CBIR concerns in biomedical disciplines,
where content-based indexing and retrieval based on information
contained in the pixel data of medical images is expected to have a
great impact on biomedical image databases [9]. However, existing
systems are not well-suited to the medical imagery special needs,
because several images out of the context are mixed with the rel-
evant ones. Thus, novel methodologies are urgently needed. An-
other use is because image collections expand very quickly today
with the price of storage going down, faster internet, and off-the-
shelf price for personal digital cameras.
It is important to recognize the shortcomings of CBIR as a real-
world technology. One problem with all current approaches is
the reliance on visual similarity for judging semantic similarity,
which may be problematic due to the semantic gap between the
low-level content automatically extracted and the higher-level con-
cepts sought by the user [6].
Basically, all systems use the assumption of equivalence of an
image and its representation in the feature space. These systems
often use measurement systems such as the easily understandable
Euclidean vector space model for measuring distances between a
query image (represented by its features) and possible results, rep-
resenting all images as feature vectors in a n-dimensional vector
space. Still, the use of high-dimensional feature spaces has shown
to cause problems. Also, caution should be taken when choosing
the distance measure in order to retrieve meaningful results. These
problems with a similarity definition in high-dimensional feature
spaces is also known as the curse of dimensionality, and has also
been discussed in the domain of medical imaging [10].
Beyer in [4] proved that the increasing in the number of features
(and consequently the dimensionality of the data) leads to losing
the significance of each feature value. Thus, to avoid decreasing
the discrimination accuracy, it is important to keep the number of
features as low as possible, establishing a trade-off between the
discrimination power and the feature vector size.
Aimed at overriding the problems of the semantic gap and the
curse of dimensionality, this paper shows a simple but power-
ful feature extractor based on multiresolution wavelet transform,
which uses the approximation subspace to compose the feature vec-
tor to represent the image. The results of applying our method
achieves 90% regarding the precision in the retrieval of medical
images that asks up to 65% of the image set.
The remainder of this paper is structured as follows. Section 2
presents the main concepts needed to follow the paper. Section 3
presents the proposed method, while Section 4 discusses the exper-
iments and results achieved with the developed method. Finally,
Section 5 presents the conclusions of this work.
2. BACKGROUND
The characteristics (features) extracted from the multimedia data
is one of the key aspects for the similarity comparison between
complex data (e.g. images, videos, sounds, time series, DNA se-
quences, among others). As afore cited, the features are arranged
in a vector named feature vector that can be seen as a n-dimensional
point in a vector space.
We introduce in the present Section a brief review about the mul-
tiresolution method and wavelet theory. We also describe the tra-
ditional histograms, as well as the distance functions that we will
in this paper to quantify the similarity between the feature vectors,
and we describe the Statistical Association Rules Miner algorithm,
called StARMiner.
2.1 Wavelets
Wavelets are mathematical functions that separate the signal in
different components of frequency, and then examine each compo-
nent with a combined resolution with its scale. The wavelets have
a multiresolution property that make it easier to extract the image
features from transformed coefficients [14].
Our proposed technique works on image subspaces generated by
applying wavelet transforms through the multiresolution method.
The central element of a multiresolution analysis is a function
(t), called the scaling function, whose role is to represent a signal
at different scales. The translations of the scaling function con-
stitute the building blocks of the representation of a signal at a
given scale. The scale can be increased by dilating (stretching) the
scaling function or decreased by contracting it [14].
The scaling function (t) acts as a sampling function (a basis),
in the sense that the inner product of (t) with a signal represents
a sort of average value of the signal over the support (extent) of .
A recursive application of this process generates new nested spaces
V j , i.e., ...V 2 V 1 V 0 V 1 ..., which are the basis of [10].
the multiresolution analysis.
By definition, a signal in V 1 can be expressed as a superposi-
tion of translations of the function 1 , but because the space V 0 is
included in V 1 , any function in V 0 can also be expanded in terms
of the translations of (t). In particular, this is true for the scaling
function itself.
Consequently, there must exist a sequence of numbers h = {h0 ,
X
h1 , . . .} such that the following relationship is satisfied :
n (or Lp norm) [17], which is employed to vector spaces. In a vec-
0 (t) = hn 1 (t ) (1)
2 tor space the objects are identified with n real-valued coordinates
n
Equation 1 is very important and it is known as the scaling equa-
tion. Equation 1 describes how the scaling function can be gener-
ated by superposing compressed copies of itself. Now it is possible
to define a new space W j as the orthogonal complement of V j in
V j+1 . In other words, W j is the space of all functions in W j that
are orthogonal to all functions in V j under the chosen inner prod-
uct. The relationship to wavelets is in the fact that the spaces W m
are spanned by dilation and translation of a function (t), thus,
such collection of basis functions are called wavelets.
As in the case with the scaling function, since the wavelet (t)
belongs to V 1 , it can be expressed as a linear combination of (t)
(t) =
X
at scale m = 1, which can be written as:
gn 1 (t n) (2)
n
where the sequence g is called the wavelet sequence. In the lit-
erature, h and g are known as the low and high frequency filters
respectively.
(a) (b) (c)
Figure 1: Example of wavelet decomposition (a) original im-
age; (b) image decomposed in two steps from Haar wavelets;
(c) configuration of regions after decomposition
The choice of a wavelet basis still represents an open problem for
filtering [7]. Probably the most popular wavelets are the Daubechies
wavelets, because of their orthogonality and compact support [16].
We choose Symlets, Coifman and Daubechies wavelets to explore
in this work.
Figure 1 shows an example of a wavelet decomposition and the
configuration of regions after decomposition.
2.2 Traditional Gray-Level Histogram
One of the most common technique used to represent an image
regarding to gray-level (color) content is the traditional histogram.
It gives the frequency of occurrences of a specific gray-level ob-
tained from the pixels of the image. Its omni-presence is mostly
due to its nice properties of linear cost to be obtained, as well as
its invariance on rotation, translation and scale, for normalized his-
tograms. It can also be used as a first step on selecting the most
relevant images for a query, thus reducing the candidate set, be-
fore applying a more costly feature extractor to compare the images
2.3 Distance Functions
The efficiency and efficacy of multimedia data retrieval will be
significantly affected by the inherent ability of the distance function
on separate the data [5]. Thus, we briefly describe some of the
distance functions that we evaluated in this paper.
2.3.1 Minkowski Family
The most widely used distance function is the Minkowski family
vuuX
{x1 ,..., xn }. Thus the Lp distances are defined as:
t
n
Lp ((x1 , ..., xn ), (y1 , ..., yn )) = p
|xi yi |p (3)
i=1
The well-known Euclidean distance corresponds to L2 . Accord-
ing to the value assigned to p we obtain the Lp family variations.
The L2 (Euclidean) distance is commonly used to calculate the
distance between vectors, and corresponds to the human-beings no-
tion of distance. This distance is additive, in the sense that each
vuuX
feature contribute independently to the measure of distance and it
is formally defined as:
dL2 (X, Y ) = t k
(xi yi )2 (4)
i=1
2.3.2 Statistic Value 2
This distance function is calculated as the difference between
each observed and theoretic frequency for each possible result, ris-

Image
Database
Figure 2: Proposed method of feature extraction using 4 levels of decomposition. Each pixel value of the approximation subspace is
assigned to the feature vector
ing it to square, and dividing each one by the theoretic frequency.
Finally, the 2 is defined as:
d2 (X, Y ) =
Xn
(yi mi )2
(5)
i=1
mi
where mi = yi +x i
. This distance function emphasizes the el-
2
evated discrepancies between two feature vectors X and Y com-
pared, and measures how improbable the distribution is.
2.4 Association Rules
In this work, associatin rules are employed to perform feature
selection and, consequently, dimensionality reduction of the fea-
ture vectors. Mining of association rules [1] is one of the most
investigated areas in data mining. Currently, finding associations
has been widely used in many applications such as customer cate-
gorization, data classification and summarization [11, 12]. Mining
images demands to extract their main features regarding specific
criteria. After extracted, the feature vector and image descriptions
are submitted to the mining process.
The process of dimensionality reduction (also called dimension
reduction) aims at reducing the number of features (attributes or
dimensions) used to represent a dataset under consideration. Thus,
to achieve the dimensionality reduction of feature vectors can be
employed an algorithm named StARMiner [13], to mine statistical
association rules from features of a training dataset, which is briefly
described in the following section.
2.4.1 Statistical Association Rule Mining
The essence of this approach is the concept of association rule
mining. Traditionally, the problem of mining association rules con-
sists to find relationships of the form A B, where A and B are
set of items, indicating that A and B frequently occurs together in
the database transactions, and that, if A occurs there is a high prob-
ably that B also occurs. These type of rules work well when deal-
ing with categorical (nominal) data items. However, when dealing
with image features that consists of continuous attributes, a type of
association rule that considers continuous values is necessary. A
recent type of continuous association rules is the statistical asso-
ciation rules, which are rules generated using statistical measure-
ments.
Let T be a dataset (for instance of medical images), xj an im-
age class, Txj T the subset of images of class xj and fi the ith
feature of the feature vector F . Let fi (Z) and fi (Z) be, respec-
tively, the mean and standard deviation of the values of feature fi in
the subset of images Z. The algorithm uses three thresholds defined
by the user: min - the minimum allowed difference between the
average of the feature fi in images from class xj and the average of
kNNqueryrequest
withtheappropriate
distancefunction
Indexing
Indexed
feature Queryresults
vectors
User
fi in the remaining dataset; max - the maximum standard devia-
tion of fi values allowed in a given class and; min - the minimum
confidence to reject the hypothesis H0. StARMiner mines rules of
the form xj fi , if the conditions given in Equations 6, 7 and 8
are satisfied.
fi (Txj ) fi (T Txj ) min (6)
fi (Txj ) max (7)
H0 : fi (Txj ) = fi (T Txj ) (8)
In Equation 8, H0 should be rejected with a confidence equal
to or greater than min , in favor of the hypothesis that the means
fi (Txj ) and f (T Txj ) are statistically different. A rule xj
fi , returned by the algorithm, relates a feature fi with a class xj ,
where values of fi have a statistically different behavior in images
of class xj . This property indicates that feature fi can distinguish
images of class xj from the remaining ones.
The StARMiner algorithm associates classes and features with
the highest power to distinguish the images. The features returned
in the rules have a particular and uniform behavior in images of a
given category. This is important because the features presenting
uniform behavior to every image in the dataset, independently of
the image category, do not contribute to categorize them and should
be eliminated from the feature vector. Therefore, the StARMiner
rules are very useful to reveal the relevance of the image features.
3. THE PROPOSED METHOD
Our method was built aimed at dealing with two inherent draw-
backs of a CBIR system: the high dimensionality of feature vectors
and the semantic gap. We amend the first one by applying higher
resolution on the multiresolution technique, and the second one by
characterizing images through the feature vectors composed of the
approximation subspace, which is obtained through a convolution
over each image by the wavelet filters. We choose the following
wavelet filters: Coifman (coif 1 and coif 2), Symlet (sym2, sym3,
sym4, sym5 and sym15) and Daubechies (db1, db2, db3, db4 and
db8). The wavelet mnemonics used in this work are the same ones
employed in the Matlab 6.5 tool.
We use 4, 5 and 6 levels of resolution and the approximation
subspace is represented by reading it column to column, assign-
ing the values obtained to the feature vector. Thus, it is important
to highlight that in our method the feature vector is composed by
the image, however in a reduced scale. Hence, this methodology
establishes a new approach. Equation 9 gives the formula to calcu-
late the number of elements of the proposed feature vector, where
width is the image width, height is the image height and n is the
level of decomposition.

width height
#f eatures = , (9)
2n 2n
For instance, if an image has 256 256 pixels and 4 levels of
resolution are applied, the vector has 256 features; when 5 levels
are applied, the vector has 64 features. And when 6 levels are used,
the vector has 16 features, which is the total of pixels from the
approximation subspace.
As different distance functions can separate better specific fea-
ture vectors, it is important to access the most suited ones to each
type of features from the image databases. Thus, we combine the
feature vector proposed with the distance function appropriate for
the calculated features.
Figure 2 graphically summarizes the proposed method and the
prototype developed, which is describe in the Section 4.
4. EXPERIMENTS AND RESULTS
In order to evaluate the effectiveness of the proposed technique
we worked on a variety of medical images categories and with the
Amsterdam Library of Object Images (ALOI) [8]. As an efficacy
measure we also have generated graphs based on the approach of
Precision vs. Recall [2], obtained from the results of sets of simi-
larity query.
Recall indicates the proportion of relevant images in the database
that has been retrieved when answering a query, and Precision is the
portion of the retrieved images that are relevant for the query. As a
rule of thumb, the closer a Precision vs. Recall curve to the top of
the graph, the better the technique. They are formally defined as:
|RA | |RA |
P recision = Recall =
|R| |A|
where, RA is the number of retrieved and relevant images; R is the
number of all images retrieved by the query; A is the number of
relevant images in the dataset.
Query Image
Thumbnails of the 10 nearest neighbors
0037.jpg
Figure 3: Example of a 10-nearest neighbor query performed
by the developed prototype over an image database of 704 im-
ages
To build the Precision vs. Recall graphs, we applied sets of k-
nearest neighbor (k-NN) queries. A k-NN query consists in search-
1 ing the dataset for the k closest images to the query image, com-
paring their feature vectors by using a defined distance function to
0.8 compute how similar the images are. Using the proposed method,
we developed a prototype to process k-NN queries. Figure 3 shows
0.6 an example of a k-NN query performed by our prototype.
For our experiments on medical image database, each Precision
vs. Recall curve represents the average curve of all the curves ob-
0.4 tained by performing a k-NN query for each image in the whole
image set, and for the ALOI database, we use just 20% of the data-
0.2 set to perform the k-NN query, due to the database size. This is
0
0 0.2 0.4 0.6 0.8 1
Recall
30 features 16 features
a common configuration employed to perform a large number of
tests over the algorithm under evaluation.
In order to accelerate the similarity queries processing, we used
the Slim-tree [15] as the indexing structure for our prototype, which
is a metric access method (MAM) specially developed to minimize
disk accesses, making the whole system faster. Furthermore, we
experiment several metrics (see Section 2.3) to perform the k-NN
queries.
4.1 Experiment 1 - The 210 Images Dataset
This dataset consists of 210 medical images classified in seven
categories: Angiogram, MR (Magnetic Resonance) Axial Pelvis,
MR Axial Head, MR Coronal Abdomen, MR Coronal Head, MR
Sagittal Head and MR Sagittal Spine. Each category is represented
by 30 images.
First, we compare the method by using 4 levels of resolution
and Coifman (coif 1 and coif 2) , Daubechies (db2 and db8) and
Symlet (sym2, sym3, sym4, sym5 and sym15) wavelets. The
use of each one of these wavelet transforms generate a vector with
256 features. The Precision vs. Recall curves of the nine proposed
vectors are shown in Figure 4. Each point of the graph is obtained
by the average of 210 queries. To perform the present experiments
we employed the L2 distance function.
Analyzing the graphs of Figure 4, we notice that the wavelet
that best represents the images is the Daubechies db2. We can also
consider that the curve generated by coif 1 wavelet practly ties to
db2.
1
0.9
Precision
(10) 0.8
0.7
0.6
0.5
0 0.2 0.4 0.6 0.8 1
Recall
coif1 db8 sym4
coif2 sym2 sym5
db2 sym3 sym15
Figure 4: Precision vs. Recall curves showing the retrieval be-
havior of the proposed method with 4 resolution levels using
256 features with the L2 metric
For the graphic of Figure 5, we compare the Coifman (coif 1 and
coif 2) and Daubechies (db1, db2 and db8) wavelet transforms in 5
resolution levels. Thus, their feature vectors also have 64 features.
Figure 5 shows the Precision vs. Recall curves from the queries on
the dataset represented by these feature vectors.
Note that the Haar (or db1) wavelet is the best one to represent
these images. As a basis for comparison, the graph in Figure 7 also
presents the average Precision vs. Recall curve obtained by using
gray-level histograms over the same image dataset. The results for
Haar (or db1) give the best Precision vs. Recall curves shown until
now. We can see that all the proposed methods have better Preci-
sion vs. Recall curves than Histogram. The curve with the highest
precision is the one generated by db1, with just 64 features, while
the other methods use 256 features, i.e., there is a reduction of 75%
 Recall Recall
coif1-4n db1-5n coif1 db8 sym4
db2-4n histogram coif2 sym2 sym5
db2 sym3 sym15
1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

Precision
Precision

0.7

Precision
0.7 0.7

0.6 L1 0.6 L1 0.6 L1

L2 L2 L2
0.5 Linf 0.5 Linf 0.5 Linf
Chi Chi Chi
0.4 Jeffrey 0.4 Jeffrey 0.4 Jeffrey
Canberra Recall Canberra Recall Canberra
Recall
0.3 0.3
0 0.2 0.4
1 0.6 0.8 1
0.3
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) (b) (c)
0.9
Figure 6: Precision vs. Recall graphs illustrating the comparison among several distance functions using Daubechies db1 wavelet

Precision
0.8 of resolution
with: (a) 4 levels of resolution; (b) 5 levels of resolution; and (c) 6 levels

1 0.7
Feature Extractor - Texture
L2
1
Feature Extractor - Texture
Linf
1 L2 Starminer
These methods are well-suited to represent the
Weighted L2 k=0
Weighted L2 k=1
images
Linf Starminer
Weighted Linf k=0
Weighted Linf k=1
under
0.8 0.6evaluation, since the precision
0.8 values are over than 80% for all re-
0.9 call values smaller than 65%. It is important to emphasize that the
0.5regions of low recall 0.6are the most important into a CBIR system
Precision (%)

Precision (%)
0.6
Feature Extractor - Texture
0.8
Precision

1
L1 because k-nearest queries usually dont search for high values of k.
L1 Starminer
Weighted L1 k=0
Weighted L1 k=1
0.4 0Thus, by0.2
comparing0.4
0.4
the results0.6provided0.8by the feature 1 vectors
0.8
0.7 with 256 and 64 elements, Recall
where the smaller one brings a more
0.2 0.2
accurate image set, we can conclude that the dimensionality curse
Precision (%)

0.6
0.6 0 coif1
really damages the results.
0
db8
This happens because sym4 the irrelevant fea-
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0.4
Recall (%) coif2
tures sym2
disturb the influence of the relevantRecall sym5
ones.
(%) Moreover, the appli-
0.5 db2of wavelet transform
cation sym3 in 5 levels through sym15 the multiresolution
0.2 method reduced the redundancy of information from data, and it
0 0.2 0.4 0.6 0.8 1 also well represents the images for executing similarity queries.
0
0 0.2 0.4 0.6 0.8 1 Recall
Recall (%)
4.2 Experiment 2 - The 704 Image Dataset
coif1 db1 db8
A larger image dataset, with 704 MR images, which is classified
coif2 db2
in eight categories was used herein. The number of the images in
Figure 5: Precision vs. Recall curves showing the retrieval be- the dataset regarding each category is: Angiogram (36), MR Axial
havior of the proposed method with 5 resolution levels using 64 Pelvis (86), MR Axial Head (155), MR Sagittal Head (258), MR
features with the L2 metric Coronal Abdomen (23), MR Sagittal Spine (59), MR Axial Ab-
domen (51) and MR Coronal Head (36). Figure 8 shows one image
from each category. As in the previous experiment the best result
in data dimensionality. The queries performed by using the db1 was obtained applying a Daubechies wavelet transform, and, ac-
feature vectors with 5 level of resolution gives precision levels up cording to Wang [16], the Daubechies wavelet achieves excellent
to 82,55% regarding the images histogram, to queries that ask up results in image processing due to its properties. Then we use sev-
to 90% of the images. eral wavelets of the family of Daubechies on 4, 5 and 6 levels of
resolution by the multiresolution method.
1

0.8
Precision

(a) (b) (c) (d) (e) (f) (g) (h)

0.6
Figure 8: Examples of images from the dataset. (a) Angiogram,
(b) MR Axial Pelvis, (c) MR Axial Head, (d) MR Axial Ab-
0.4
domen, (e) MR Coronal Abdomen, (f) MR Coronal Head, (g)
MR Sagittal Head, and (h) MR Sagittal Spine
0.2
We evaluated several distance functions such as L1 , L2 , L ,
0 Jeffrey Divergence, Canberra and 2 . The L1 , L2 and 2 metrics
0 0.2 0.4 0.6 0.8 1
presented the best results. We can observe that the curves of Preci-
Recall sion vs. Recall obtained by these distance functions are equivalent,
as we can see in Figure 6. Due to this fact we employed the Eu-
coif1-4n db1-5n clidean distance function to perform the experiments.
db2-4n histogram Figure 9 shows the Precision vs. Recall curves generated by the
Figure 7: Precision vs. Recall curves showing the retrieval be- proposed method. First, the wavelet name is displayed, then the
havior of the best curves and gray-level histograms with the L2 level of resolution and finally the number of elements of the feature
metric vector. Observe that the Precision vs. Recall curves generated by
.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Recall Recall
db8 coif1-4n db1-5n coif1 db8 sym4
db2-4n histogram coif2 sym2 sym5
db21 sym3 sym15
the same wavelet transform in several levels of resolution decrease
according to the wavelet chosen. The bigger the number of the fil-
ters, the fast the curves decrease when a higher level of resolution is 0.8
chosen. Analyzing the db1 wavelet, which has two filters, note that

Precision
the curves generated in 4 and 5 levels of resolution are equivalent, 0.6
even exiting a large difference between the number of features from
their vectors. For a 6 levels of resolution we still have an excellent
result (see db1-6n-16 curve), as with just 16 features, the precision 0.4
is over than 80% for values of recall up to 90%. And comparing
the 256 features with the 16 ones, we have a dimensionality reduc- 0.2
tion of 93.75%. Also, note that the larger the number of filters, the
smaller the precision of the queries, considering the same level of 0
resolution. These results suggest that relevant features are removed 0 0.2 0.4 0.6 0.8 1
from the image when we use filters that are greater than 2.
Recall
1 db1-4n db1-6n
db1-5n histogram
0.8
Figure 10: Precision vs. Recall graphs showing the retrieval
behavior of the best curves and gray-level histogram with L2
Precision

0.6 metric

0.4 D1 x01 y01 m1 1 b1 ... DL x0L y0L mL L bL

0.2
Featuresofthetextureclass1 FeaturesofthetextureclassL

0 Figure 11: The feature vector

0 0.2 0.4 0.6 0.8 1
Recall
db1-4n-256 db3-4n-256 a reduced feature vector, that in this case has selected 16 features.
db1-5n-64 db3-5n-64 Thereafter the reduced feature vectors were used to index the im-
db1-6n-16 db4-4n-256 ages in a CBIR environment and to1answer similarity queries.
db2-4n-256 db4-5n-64 Figure 12 shows the comparison of the curves generated by our
db2-5n-64 db8-4n-256 method with 16 features (db1-6n-16 curve) to the method that uses
db2-6n-16 db8-5n-64 an improved version of EM/MPM0.8 algorithm and to the method that
uses the data mining algorithm StARMiner over the feature vec-
Precision

Figure 9: Precision vs. Recall graphs generated by several
Daubechies wavelets transforms in 4, 5 and 6 levels of multires-
olution with L2 metric
Figure 10 shows the best curves of Precision vs. Recall with 256,
64 and 16 features, respectively, from the Figure 9 and compare
them with the curve given by the gray-level histogram. It is clear
to see that all three methods have a better performance than the
histogram. Numerically, we get an improvement of precision up
to 531% to values of recall up to 95%, for the feature vector with
256 features. For 64 features, the improvement in precision is up
to 528% to a recall of 95%; and for 16 features, the improvement
1 of 90%.
in precision is up to 491% to a recall
To compare this method with another one from literature, we
used a technique proposed by Balan 0.8 [3], which employs an im-
proved version of the EM/MPM method to segment images, and
for each region segmented based on texture, six features were ex-
Precision
0.6 of EM/MPM, which is one
tors extracted by the improved version
of the best method in the literature. We can see that our method
performs better when processing0.4 similarity queries (k-NN). Note
also that, our method demands fewer features than the EM/MPM.
We can also compare the time spending to process the image and
to extract the features. While the0.2improved version of EM/MPM
spends around 17.05 seconds per image, our method with 16 fea-
0
tures spends around 0.77 seconds. Even selecting the most relevant
features by using StARMiner, our method0 0.2 better0.4
performs than the 0.6
improved version of EM/MPM with 16 features. Recall
4.3 Experiment 3 - The ALOI Image Dataset
30 features 16 fea
The third image dataset employed in the experiments is the Am-
sterdam Library of Object Images (ALOI) [8], a gray image collec-
tion of one-thousand small objects, recorded for scientific purposes
in several configurations. The ALOI-ILL dataset consists of 24,000
images in gray-scale recorded under varying 24 different illumina-

tracted: the mass (m); the centroid0.6(xo and yo); the average gray tion angle of each object. The images are represented by 8 bits,
level (), the Fractal dimension (D); and the linear coefficient used resulting in 256 gray-levels and comprising dimensions of 384 x
to estimate D (b). Therefore, when 0.4 an image is segmented in L 288 pixels. Figure 13 shows a sample of 10 elements of this data-
classes, the feature vector has L * 6 elements. Here we use L = 5, set. As in the previous experiments the best results were obtained
so the feature vector has 30 features.
0.2 Figure 11 illustrates the fea- using the db1 wavelet, we applied our method with the same one.
ture vector described. In this experiment, the Precision vs. Recall curves are built as
Over these image feature vectors was executed the data mining the average plots of 240-nearest neighbor queries, using 20% of
0
algorithm StARMiner, which was described in Section 2.4.1. The the dataset. To evaluate the correctness of each image retrieved, we
0.4 0.6 0.8 1
rules mined identify the most relevant 0 image features,
0.2 0.4
generating 0.6 used 0.8 1
the object portrayed in the image as a class attribute, therefore
Recall Recall

db1-6n 30 features 16 features

histogram
 0.6

Pr
Precision
0.4

0.4
0.2

0.2
1
1 0
0
0
0 0.2 0.4 0.6 0.8 1 0.8
0.8 Recall
5n - 108 features 6n - 30 features
0.6

Precision
0.6 5n - 60 features SM 6n - 20 features SM
Precision

0.4
0.4
1
0.2 0.9
0.2
0.8 1 30 features (EM/MPM)

Precision
15 features (StARMiner) 0.8
16 features (db16n) 0
56 0 0 0.2 0.4 0.6 0.8 0.7 1
0 0.2 0.4 0.6 0.8 1 Recall
64
0.6
56 Recall
64 6n - 27 features 5n - 108 features Histogram
0.5the ALOI-
Figure 15: Precision vs. Recall graphs obtained over
56 Figure 12: Precision vs. Recall graphs from db1-6n with 16
64 features and from improved version of the EM/MPM ILL dataset for each category employing: db1 with0 5 and0.26 0.4
1 levels of decomposition with 2 and the StARMiner feature se- Rec
lection
coif1 db8
0.8 coif2 sym2
db2
built by executing similarity queries employing: (1) the proposed sym3
0.6 feature vector comprising 27 features, using db1 with 6 levels of
decomposition; (2) the proposed feature vector comprising 108 fea-
0.4 tures, using db1 with 5 levels of decomposition and (3) the features
obtained by using the traditional gray-level histograms, each vector
0.2 comprising 256 features, since the images are represented by 8 bits.
Figure 13: Examples of images from the dataset Thus, analysing the graphs of Figure 15, we can notice that both
0 instances of the proposed feature vectors (with 27 features and
0 0.2 0.4 0.6 0.8 1
108 features) presented a considerable improvement in precision,
performing a supervised automated evaluation of the Recall algorithm. up to 75% and 125% for a recall level of 45% and 50% respec-
For each query, the images considered for the computation of pre- tively in comparison with the precision achieved by the traditional
cision are those related to the same classdb1-4n
in the dataset. The plots db1-6n gray-level histogram. Hence, these results testify that the proposed
db1-5n histogram
show the average result for all images of the dataset as query im- feature vector improves the precision of similarity queries. In such
ages. case, we can argue that our proposed feature vector is well-suited to
We evaluated the L1 , L2 , L , Jeffrey Divergence, Canberra and content-based medical images, as well as to a generic image dataset
2 distance functions. The best results were obtained using the 2 like the ALOI-ILL. In addition to the improvement of precision, it
distance function, as we can notice in Figure 14. is important to highlight also that the method performs a notable
dimensionality reduction in comparison with the dimensionality of
1
the traditional gray-level histograms (e.g. 256 dimensions), up to
90% and 58% considering respectively the db1 with 6 levels of de-
0.8 composition (i.e. Figure 15 curve 6n-27 features) and with 5 levels
(i.e. Figure 15 curve 5n-108 features).
0.6 In Figure 16 is illustrated the Precision vs. Recall graphs ob-
Precision

tained by using the proposed feature vector comprising 27 features

0.4
(i.e. db1 with 6 levels of decomposition), 108 features (db1 with
L1 5 levels of decomposition) and also the features selected by the
L2
Linf StARMiner algorithm. When analysing these graphs we can ob-
0.2 Chi serve that again the proposed feature vector clearly improves the
Jeffrey precision of similarity queries, but it is remarkable to note that the
Canberra
0
0 0.2 0.4 0.6 0.8 1
precision obtained by the StARMiner almost ties with the preci-
Recall sion obtained by our original feature vector. It is important to high-
light that although the precisions tie, combining the StARMiner to
Figure 14: Precision vs. Recall graphs illustrating the com-
our proposed feature vector provides a dimensionality reduction of
parison among several distance functions using Daubechies db1
aproximately 26% and 45% on the feature vectors size, considering
wavelet with 6 levels of resolution
the vectors that comprise 27 and 108 features respectively. Hence,
demanding less space and making the processing faster, and at the
The Precision vs. Recall graphs of Figures 15 and 16 correspond
same time preserving the precision of the similarity queries.
to the experiments performed on the ALOI-ILL image dataset rep-
resented by the proposed feature vector described in Section 3. The
results were obtained using the 2 distance function since it demon- 5. CONCLUSIONS
strates the better precision gain in comparison with the traditional In this paper we presented a technique based on wavelet approx-
ones (e.g. Minkowski family). imation subspaces, which was used to compose the image feature
The Precision vs. Recall curves in the graphs of Figure 15 were vector to process similarity queries on the image content. A tool

0.8
cision

0.6
 1 [2] R. A. Baeza-Yates and B. A. Ribeiro-Neto. Modern
1
0.8
s [3] A. G. R. Balan, A. J. M. Traina, C. Traina Jr., and P. M. d. A.
M Marques. Fractal analysis of image textures for indexing and
Precision
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 Springer Verlag.
Recall
6n - 30 features 5n - 108 features
6n - 20 features StaRMiner 5n - 60 features StaRMiner
Figure 16: Precision vs. Recall graphs obtained over the ALOI-
ILL dataset for each category employing: db1 with 5 and 6
levels of resolution with 2 metric and the StARMiner feature
selection
based on the presented technique was implemented, first aiming
at validating the technique proposed on real images from different
tissues of the human body, and to assist the study and analysis of
medical images, aiming to be included in a PACS under develop-
0.4 0.6 ment 0.8in our 1institution; second aiming at validating it on a generic
Recall
image dataset, that serves to scientific proposes.
First, several wavelets were evaluated to the medical image data-
db1-6n
set, and the Daubechies showed a better efficacy than the other
histogram
ones for the analyzed image sets. The achieved results showed
that the proposed method performs very well, presenting an image
retrieval accuracy always over 90% for recall values smaller than
65%. Moreover, we obtained a feature vector with just 16 elements
that provided a better performance than the vector with 30 features
obtained from segmented images by using an improvement version
of EM/MPM algorithm, which is a much more time consuming
method.
Second, as in the previously experiment we got better results ap-
plying the filters of Daubechies, we applied our method using these
filters on the ALOI-ILL, which is a generic image database and
larger than the other one. We could notice that the curves of Pre-
cision vs. Recall were not equivalents. We had best results apply-
ing our technique using four levels of decomposition. In addition,
combining the proposed method with StARMiner we improved the
queries precision. Thus, it can be inferred that for this database
our method generated a feature vector that still had some redun-
dant features when was used five levels of resolution (108 features).
Then, we moved on to reduce the dimensionality of the feature vec-
tor using six levels of resolution (27 features). But, this procedure
leaded to lose some significant features. That is why our curves of
Precision vs. Recall decreased. In this case, the combination of
our method and StARMiner provided a better representation of the
images.
By the results obtained in the work, we can claim the wavelets
and the multiresolution method are well suited to deal with the issue
of the semantic gap and the dimensionality of feature vectors.
6. REFERENCES
[1] R. Agrawal and R. Srikant. Fast algorithms for mining
association rules. In International Conference on Very Large
Databases (VLDB), pages 487499, Santiago de Chile,
Chile, 1994.
Information Retrieval. Addison-Wesley, Wokingham, UK,
1999.
retrieval by content. In 18th IEEE Intl. Symposium on
Computer-Based Medical Systems - CBMS, pages 581586,
Dublin, Ireland, 2005. IEEE Computer Society.
[4] K. Beyer, J. Godstein, R. Ramakrishnan, and U. Shaft. When
is nearest neighbor meaningful? In C. Beeri and
P. Buneman, editors, ICDT, volume 1540 of Lecture Notes in
Computer Science, pages 217235, Jerusalem, Israel, 1999.
[5] P. H. Buggatti, A. J. M. Traina, and C. T. Jr. Assessing the
best integration between distance-function and image-feature
to answer similarity queries. In 23rd Annual ACM
Symposium on Applied Computing, pages 12251230,
Fortaleza, CE, Brazil, 2008.
[6] T. M. Deserno, S. Antani, and L. R. Long. Gaps in
content-based image retrieval. In S. C. Horii and K. P.
Andriole, editors, Medical Imaging 2007: PACS and
Imaging Informatics, volume 6516 of SPIE, 2007.
[7] C. Garcia, G. Zikos, and G. Tziritas. Wavelet packet analysis
for face recognition. Image and Vision Computing,
18(4):289297, 2000.
[8] J.-M. Geusebroek, G. J. Burghouts, and A. W. M. Smeulders.
The Amsterdam library of object images. Journal of
Computer Vision, 61(1):103112, 2005.
[9] S. K. Kinoshita, P. M. d. A. Marques, R. R. P. Jr., J. A. H.
Rodrigues, and R. M. Rangayyan. Content-based retrieval of
mammograms using visual features related to breast density
patterns. Journal of Digital Imaging, 20(2):172190, 2007.
[10] H. Mller, N. Michoux, D. Bandon, and A. Geissbuhler. A
review of content-based image retrieval systems in medical
applications-clinical benefits and future directions.
International Journal of Medical Informatics, 73(1):123,
2004.
[11] C. Ordonez, N. Ezquerra, and C. A. Santana. Constraining
and summarizing association rules in medical data.
Knowledge and Information Systems, 9(3):259283, 2006.
[12] H. Pan, J. Li, and Z. Wei. Mining interesting association rules
in medical images. In Advance Data Mining and Medical
Applications, pages 598 609. Springer Verlag, 2005.
[13] M. X. Ribeiro, A. G. R. Balan, J. C. Felipe, A. J. M. Traina,
and C. Traina Jr. Mining statistical association rules to select
the most relevant medical image features. In IEEE MCD05,
pages 9198, Houston, USA, 2005. IEEE Computer Society.
[14] E. J. Stollnitz, T. D. DeRose, and D. H. Salesin. Wavelets for
Computer Graphics - Theory and Applications. Morgan
Kaufmann Publishers, Inc, San Francisco, CA, 1996.
[15] C. Traina Jr., A. J. M. Traina, B. Seeger, and C. Faloutsos.
Slim-trees: High performance metric trees miniminzing
overlap between nodes. Research Paper CMU-CS-99-170,
Carnegie Mellon University - School of Computer Science,
October 1999 1999.
[16] J. Z. Wang. Wavelets and imaging informatics: A review of
the literature. Journal of Biomedical Informatics,
34:129141, 2001.
[17] D. R. Wilson and T. R. Martinez. Improved heterogeneous
distance functions. Journal of Artificial Intelligence
Research, 6:134, 1997.