INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

Texture Image Retrieval Using Fourier

Transform
Abdelhamid Abdesselam1,

simpler and faster. As a consequence, wavelet-based

Abstract— Texture is an important visual property that approaches gained much more popularity among the
characterizes a wide range of natural and artificial images computer vision community. Fourier transform has also
which makes it a useful feature for retrieving images. During been widely used in characterizing textures. One of the
the last decades, several approaches have been proposed to reasons for that is its suitability for describing periodic
describe the texture contents of an image. In early research functions such as texture images which usually contain
works, texture descriptors were mainly extracted from the
pixel space itself, edge histograms and co-occurrence-based
quasi-repetitive patterns. Concentrations of Fourier power
features, are examples of such descriptors. Later on, dual spectrum values capture dominant orientations of the
spaces (transform of pixel space) such as frequency space or patterns in the image and their distribution in the frequency
spaces resulting from Gabor and wavelet transforms were space is closely related to coarseness of the texture [5, 6].
explored for texture characterization. This paper describes a These two features (directionality and coarseness of a
Fourier-based technique for characterizing image textures. texture) are of importance in texture analysis [17]. The main
The performance of the technique is compared with several drawback of using Fourier transform is the poor spatial
Fourier- and wavelet-based methods described in literature. localization it provides. Windowed Fourier transform has
The two main criteria that were used in this comparison are been introduced to overcome this problem at the cost of a
the accuracy and execution time of the techniques.
significant increase in the computations [18]. In this paper
we describe a Fourier-based method for characterizing
Index Terms— Fourier Transform, Texture-based Image
Retrieval, Wavelet Transform. texture images and show that it performs better than many
of the recently proposed techniques for texture
I. INTRODUCTION characterization. We have included in our comparative
study several Fourier- and wavelet-based texture
Content-Based Image Retrieval (CBIR) has been an characterization techniques. The two main criteria that were
active research topic in the last two decades. As a used in this comparison are the accuracy and running time
consequence, several experimental and commercial image of the techniques.
retrieval systems have been proposed during this period of
time. In a CBIR system, image databases are queried using II. SUMMARY OF THE TECHNIQUES CONSIDERED IN THIS
the visual content of an image which is usually represented STUDY
by low level features such as color, texture, shapes, or a
combination of some or all of these features. A. Wavelet-based techniques
Texture is widely used in CBIR systems since it plays a
Wavelet-based approaches gained an increasing interest
crucial role in characterizing real world images both natural
from the computer vision community partly, because several
(such as images of clouds, water, trees, remotely sensed data
physiological studies of the visual cortex suggested the use
and medical images) and man-made ones (such as images of
of multi-scale analysis of the visual information by Visual
bricks, fabric, and buildings ).
Systems of primates [16]. Besides, wavelet transform has
Several approaches have been proposed to describe
also the very useful property of space-frequency
texture contents of an image. In early works, the texture
localization. A significant number of wavelet-based
features were mainly extracted from the pixel space itself
methods have been described in literature [8-12], most of
using edge density, edge histograms, and co-occurrence-
them use pyramidal wavelet decomposition in which the low
based–features to characterize the image texture [1-4]. More
frequency sub-band is repeatedly decomposed until a
recently, various transforms have been used to produce dual
required level is reached. Usually either means or standard
spaces from which texture features were extracted. The most
deviations or both of the energy of the sub-bands are used as
common transforms are Fourier [5-7], wavelet [8-12] and
texture features.
Gabor transforms [13-15]. While the use of Gabor filters is
The mean of the energy of a sub-band is defined as
supported by physiological evidences [16], it suffers from
shown in equation (1)
serious drawbacks such as the need to tune the parameters
of the filter and the complexity of the calculations involved
N −1
1
[15]. In the other hand, wavelet-based approaches are much
μ Energy =
N
∑w
k =0
2
k (1)

1
Department of Computer Science, College of Science, Sultan Qaboos
University, P.O. 36, Al-Khodh, Muscat-123, Sultanate of Oman; e-mail:
[email protected] where wk is the wavelet coefficient at pixel k of the sub-
band and N is the number of pixels in the sub-band.

© SQU-2009 ISSN: 1813-419X 343

INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

Sometimes the absolute value of wk ( |wk| )is used instead of The spectral domain is partitioned into half-ring-shaped
its square (wk2) just to speed up the calculations. regions and wedge-shaped regions where:
a half-ring-shaped region Rr1,r2 is defined as follows:
The standard deviation is estimated as described by the
following equation (2) Rr1, r 2 = {(u , v) | r12 ≤ u 2 + v 2 < r 22 ; n / 2 + 1 ≤ u ≤ n;0 ≤ v ≤ n}

1 N −1 2
∑
A wedge-shaped region is defined as follows
σ Enerfy = ( wk − μ ) 2 (2)
N − 1 k =0
Wθ 1,θ 2 = {(u , v) | θ 1 ≤ tan −1(v / u ) < θ 2; n / 2 + 1 ≤ u ≤ n;0 ≤ v ≤ n}
Similarly, absolute value is sometimes used instead of the
square operation.
Very recently, P.W.Huang et al [10-12] proposed a The intersections of these rings and wedges define sectors
wavelet based approach that describes texture content of an Sr1,r2,θ1,θ2, where
image by concatenating the gradient vectors of the sub-band
images obtaining a single feature vector called Composite S r1, r 2,θ 1,θ 2 = {(u , v) | r12 ≤ u 2 + v 2 < r 22 ; and θ 1 ≤ tan −1(v / u ) < θ 2,
Sub-band Gradient (CSG) vector. A gradient vector of an
image is the histogram that records the total gradient and n / 2 + 1 ≤ u ≤ n;0 ≤ v ≤ n}
magnitude of the image pixels at different directions. Their
experiments showed that their approach outperforms the In order to reduce the effect of noise on the final result,
single resolution technique using the same texture feature. only significant values of the power spectrum were
considered for further processing.
The mean and standard deviation of the significant values
B. Fourier-based techniques of power spectrum in these sectors constitute the elements
Fourier transform has been widely used by image of the feature vector.
processing research community. It has the very useful Such a feature vector describes three aspects of the
characteristic of highlighting the dominant spatial texture: its roughness, coarseness and directionality. Since
frequencies as well as the dominant orientations of the rough texture tend to contain more energy in the high
structures contained in the image. Another advantage of frequency components than smoother ones.[7], and
using Fourier transform is the fact that frequency-domain magnitude of power spectra for frequency components away
features are generally less sensitive to noise than spatial from the origin drop rapidly to approximately zero for
domain ones [5] coarse textures [5] and it's well known that linear structures
In many works described in literature [5-7], the spectral in the spatial domain at direction θ will produce linear
domain is partitioned into ring-shaped and wedge shaped structures in the frequency domain with high energy values
areas. The mean and or variances of the Fourier power at direction θ+90°.
spectrum of ring-shaped regions were used to describe the
coarseness of the texture and the variances of the wedge- A. Algorithm
shaped regions were used to describe the orientation of the The feature extraction algorithm used in our method can
textural patterns. be summarized as follows:

1- Apply Fast Fourier transform to

III. PROPOSED METHOD original image f(x,y) to get F(u,v)
2- Consider only pixels with significant
The proposed method is partly inspired from previous power spectrum values
energy-based works [5,6]. Like those approaches, we have (i.e. F(u,v)=0 if F(u,v)< Threshold)
partitioned the frequency domain into several regions. 3- Partition the Fourier image into
Unlike them, we used sectors instead of rings and wedges, sectors and calculate means and
where a sector is the intersection of a wedge and a ring see standard deviations of these sectors.
Fig.1. The advantage of using sectors is the fact that we 4- The feature vector consists of means
obtain a more accurate description of the power spectrum and standard deviations of the
distribution in the frequency domain. Each sector different sectors.
characterizes a range of orientations and coarseness of the B. Similarity Measurement
texture. Moreover, unlike the above mentioned works, we Given two k-dimensional feature vectors f1 and f2,
took advantage of the symmetry property of Fourier describing two images Im1 and Im2, the dissimilarity
transform and limited the analysis to only half of the Fourier between Im1 and Im2 is estimated using the Caneberra
space which allows us to reduce the overall processing time. distance which is defined as follows (3):
Let {f(x,y), x=1,n , y=1,n} be the texture image and
{F(u,v), u=1,n, v=1,n} be its Fourier transform. First, the
1 k f1 (i) − f 2 (i)
origin of the transformed image is shifted to the center of
the image at position (n/2+1, n/2+1). This will produce a
dissimilarity(Im1 , Im2 ) = ∑
k i=1 f1 (i) + f 2 (i)
(3)
symmetrical Fourier image. Because of this symmetry, only
half of the image is processed to extract texture features. We This metric has the advantage of reducing the impact of
have used the right half of the image defined by: the high magnitude feature components on the estimation of
{ F(u,v), n/2+1≤ u ≤ n; - n /2 ≤ v ≤ n /2-1}. the dissimilarity.

© SQU-2009 ISSN: 1813-419X 344

INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

IV. EXPERIMENT n is the number of similar images retrieved, N is the total

number of similar images in the database, and T is the
A. Parameter setting: number of short-listed images.
In our experimentation we have used twelve sectors In the experimentation we have conducted N=4, and
Sr1,r2,θ1,θ2 resulting from the intersection of three rings and T=10.
four wedges as shown in Fig.1, where We have also recorded the average time needed to
{(r1,r2)}={(0,N/6),(N/6,N/3),(N/3,N/2)}and perform feature extraction and comparison of two images
{(θ1,θ2)}={(-22.5°,22.5°),(22.5°,67.5°),(-22.5°,-67.5°), for each method tested. This allows us to compare the
(67.5°, 90°), (-67.5°, -90°)}. execution time of the various methods.
This choice is a tradeoff between having a small-sized In the experiment, we compared the results obtained by
feature vector and using enough information to discriminate our method with four other methods: two Fourier-based
different textures. The three half-ring-shaped regions ones and two wavelet based ones.
segment the Fourier domain into low, medium and high Both of the Fourier-based methods split the Fourier space
frequencies whose power spectrums are strongly related to into wedges and rings. The first one uses standard deviation
roughness and coarseness of the texture. In the other hand, of the power spectrum of the resulting regions as texture
the power spectrum of the four wedge-shaped regions are feature. The second one uses both the mean and standard
closely related to the amount of the linear structures in the deviations as texture feature. The first wavelet-based
four directions: horizontal, vertical and the two diagonals. method performs a 3-level pyramidal wavelet-
The value of the threshold defining significant power decomposition and uses energy of the sub-bands as feature
spectrum values is set to the mean of the Fourier transform descriptor similar to the one proposed by [20]. The second
coefficients. The experiments showed that such value gives wavelet-based method uses CSG feature vector as described
good results with all the images we have tested. by [12].

V. RESULTS AND DISCUSSION

We have conducted two sets of testing. The first
experiment consists of comparing the proposed Fourier
9
method with other Fourier-based ones. The second
10
experiment consists of comparing the new method with
5
6
some wavelet based methods recently proposed in literature.
1
2
Table.1 summarizes the results of the first experiment.
3
4 11
7 11 We can see that our method outperforms the traditional
1
8
Fourier method (there is about 10% improvement). This can
5
12
be explained by the fact that the splitting frequency domain
9
into sectors describes more accurately the distribution of the
power spectrum in the Fourier space than when using rings
and wedges. We can notice also that the calculations
required to build and compare the twelve-element-vectors of
the new method does not increase significantly the
Fig.1: Frequency domain is partitioned into 12 sectors execution time since it’s in the same order of the two other
each representing a specific direction and frequency range Fourier techniques (i.e. 0.01 second)

B. Test dataset CFS 72 0.01

CFM&S 72 0.01
Our dataset is made of the whole digitized Brodatz
images we downloaded from: New technique. 82 0.01
[http://www.ux.uis.no/~tranden/brodatz.html]. CFS= classic Fourier method using standard deviation
It contains 111 gray scale images of size 640 x 640 CFM&S classic Fourier method using mean and standard deviation
pixels. We have partitioned each image into 8x8 sub-images Table.1: Comparison our technique with two classical Fourier methods.
from which we have randomly selected 4 sub-images to
constitute the image database (444 sub-images) and one to Table.2 summarizes the results of the second experiment
be used as a query image. where our method is compared to two other wavelet-based
methods. We can see that in the contrary of what was
C. Performance Evaluation
reported by some researchers [21] Fourier based technique
To evaluate the performance of our approach, we have can outperform wavelet-based ones with no significant
adopted the well-known efficacy formula (4) introduced by increase in the processing time (both require 0.01s) in this
Kankahalli et al. [19] case our technique performed about 7% better than wavelet-
based technique.
⎧n / N if N ≤T Fig.2 and Fig.3 show the first 10 retrievals for query sub-
Efficacy = η T = ⎨ (4)
⎩n / T if N >T images D1-4 and D3-1 obtained by our method and the four
methods included in the comparison study
Where

© SQU-2009 ISSN: 1813-419X 345

INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

[7] D. Gibson and P.A. Gaydecki, Definition and

Technique Efficacy Time application of a Fourier domain texture measure:
(%) (seconds) Application to histological image segmentation, Comp. Biol.
WE 75 0.01 25 (1995) 551-557
CSG-Haar 70 0.04 [8] J.R. Smith and S-F. Transform features for texture
CSG-DB4 53 0.05 classification and discrimination in large image databases,
Our technique. 82 0.01 Image processing proceedings, ICIP-94, 3 407-411.
[9] M. Kokare, P.K. Biswas, B.N. Chatterji, Texture image
WE= Wavelet-based method using energy of the sub-bands as texture retrieval using rotated wavelet filters, Pattern Recognition
feature. It’s a 3-level wavelet using Daubechies-4 Letters, (2007) 28 1240-1249
CSG-Haar: CSG method using 3-level Haar-based wavelet decomposition [10] P.W. Huang, S.K. Dai, Image retrieval by texture
with36 bins(k=10)
CSG-DB4: CSG method using 3-level DB4-based wavelet decomposition similarity, Pattern Recognition, 36 (2003) 665-679.
with 36 bins (k=10) [11] P.W. Huang, S.K. Dai, Design of a two-stage content-
Table.2: Comparison our technique with two wavelet-based methods. based image retrieval system using texture similarity,
Information Processing and Management 40 (2004) 81-96.
[12] P.W. Huang, S.K. Dai, and P.L.Lin, Texture image
VI. CONCLUSION retrieval and image segmentation using composite sub-band
This paper describes a new Fourier-based method for gradient vectors, J. Vis. Communication and Image
retrieving texture images. The experiment we had conducted Representation 17 (2006) 947-957
showed that the new method outperforms several existing [13] J.G. Daugman and D.M.Kammen, Image statistics.
Fourier-based and wavelet-based methods. We are now gases, and visual neural primitives, Proc. Of IEEE ICNN 4
working on the following directions: (1987) 163-175
1. Implementing a multi-resolution version of our method [14] A.K. Jain and F. Farrokhnia, unsupervised texture
2. Studying required modifications needed to make the segmentation using Gabor filters, Pattern Recognition, 24
method rotation-invariant and (1991) 1167-1186
3. Exploring ways to perform image segmentation based [15] F. Bianconi, and A. Fernandez, Evaluation of the
on the Fourier features used by our method. effects of Gabor filter parameters on texture classification,
Pattern Recognition, 40 (2007) 3325-3335
VII. REFERENCES [16] Beck et al, Spatial frequency channels and perceptual
grouping in texture segregation, Comp. Vision Graphics and
[1] R.M. Haralick, K. Shanmugam and J. Dinstein, Textural Image Processing, 37 (1987) 299-325
features for image classification, IEEE Trans. Systems, Man [17] F.W. Campbell, and J G. Robson, Application of
and Cybernetics, 3 (1973) 610-621. Fourier analysis to the visibility of gratings, Journal of
[2] R.W. Conners and C.A. Harlow, A theoretical Physiology, 197 (1968) 551-556
comparison of texture algorithms, IEEE Trans. Pattern [18] H. Yu, M. Li, H-J Zhang, and J. Feng, Color texture
Analysis and Machine Intelligence (PAMI), 2 (1980) 204- moments for Content-Based Image Retrieval, International
222. Conference on Image Processing Proceedings ICIP02, 24-
[3] M. Amadasun and R. King , Textural features 28
corresponding to textural properties, IEEE SMC, 19 (1989) [19] M. Kankahalli, B.M. Mehtre, J.K. Wu, Cluster-based
1264-1274. color matching for image retrieval, Pattern Recognition, 29
[4] S.R. Fountain, and T.N. Tan, Efficient rotation invariant (1996)701-708
texture features for content-based image retrieval, Pattern [20] G. V. Wouwer, P. Scheunders, and D. V. Dyck,
Recognition, 31 (1998) 1725-1732. Statistical texture characterization from discrete wavelet
[5] D-M. Tsai and C-F Tseng, Surface roughness [6] J.S. representations, IEEE Trans. On Image Processing, 8
classification for castings, Pattern Recognition 32 (1999) (1999) 592-598.
389-405 [21] C-C Chen and C-C Chen, Filtering methods for texture
[6] Weszka, C.R. Dyer, A. Rosenfeld, A comparative study discrimination, Pattern Recognition Letters, 20 (1999)
of texture measures for terrain classification, IEEE Trans. 783-790.
System, man and Cybernetics, 6 (1976) 269-285.

© SQU-2009 ISSN: 1813-419X 346

INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

Our technique CFS WE CSG-Haar CSG-DB4

D1-2 D1-2 D1-2 D1-6 D1-2

D1-3 D1-3 D1-6 D1-2 D1-3

D1-6 D1-6 D1-3 D1-1 D1-6

D1-1 D95-10 D1-1 D1-3 D1-1

D95-1 D1-1 D95-10 D64-8 D64-1

D1-4
D95-10 D64-1 D64-1 D95-2 D95-1

D95-2 D94-8 D64-2 D64-2 D95-10

D64-10 D64-10 D95-2 D64-1 D95-2

D64-1 D94-6 D23-8 D95-3 D64-10

D95-3 D95-3 D95-3 D95-1 D64-8

Fig.2: First 10 retrieved images for query image D1-4 using our method and 4 other tested methods.

© SQU-2009 ISSN: 1813-419X 347

INTERNATIONAL CONFERENCE ON COMMUNICATION, COMPUTER AND POWER (ICCCP'09) MUSCAT, FEBRUARY 15-18, 2009

Our technique CFS WE CSG-Haar CSG-DB4

D3-2 D3-2 D3-2 D3-2 D3-6

D3-6 D3-10 D3-10 D3-10 D3-10

D3-10 D3-6 D3-6 D92-10 D57-6

D26-10 D85-3 D4-8 D9-1 D57-10

D84-2 D85-8 D4-2 D57-10 D4-1

D3-1

D3-3 D3-3 D97-8 D28-6 D4-3

D84-10 D82-10 D4-1 D14-2 D57-1

D102-3 D82-2 D57-6 D73-1 D2-6

D26-2 D17-2 D57-2 D71-8 D14-1

D94-10 D85-1 D4-3 D28-1 D3-2

Fig.3: First 10 retrieved images for query image D3-1 using our method and 4 other tested methods.

© SQU-2009 ISSN: 1813-419X 348