Texture Retrieval with Descriptors Based on Local Fourier

Transform: Comparing the Rectangular and Circular

Neighbourhoods
AHSAN AHMAD URSANI*, WAJIHA SHAH**, AND SYED ASIF ALI SHAH***

RECEIVED ON 01.10.2010 ACCEPTED 03.01.2011

ABSTRACT
The texture descriptors derived from 1-D DFT (Discrete Fourier Transform) of the
pixel values of a local neighbourhood have been shown to perform better than the
methods based on wavelets for image retrieval and recognition. These DFT-based texture
descriptors were extracted from rectangular or circular neighbourhoods. This paper
compares the texture descriptors extracted from rectangular and the circular
neighbourhoods previously proposed in the literature. A database of images is
constructed from Brodatz album and the texture descriptors extracted from the two
types of neighbourhoods are compared for texture retrieval. This paper shows that
extracting DFT-based features from circular neighbourhood is almost thrice as
expensive as extracting the same from the rectangular neighbourhood. The results of
image retrieval on a large image database show that the descriptor extracted from
rectangular neighbourhoods performs better than the same extracted from the circular
neighbourhoods.

Key Words: Texture Description, DFT-Based Texture Features, Circular

Neighbourhood, Rectangular Neighbourhoods, Content-Based Image
Retrieval.

1. INTRODUCTION

T
exture is an important image-content processed or retrieval. This paper concerns with texture analysis for
in Computer graphics, image retrieval, artificial the sake of retrieval.
vision systems, bio-medical imaging, land-use
classification in remote sensing and many more Varma and Zisserman [1] divided the texture features into
applications. There are two different ways in which the two broad categories, i.e. ones which put emphasis on the
texture is processed, i.e. texture analysis and texture local properties of the texture-image and others that are
synthesis. The texture is analysed either for recognition based on large filter banks. The first category included
* Associate Professor, Department of Telecommunication Engineering, Mehran University of Engineering and
Technology, Jamshoro.
** Assistant Professor, Department of Electronic & Biomedical Engineering, Mehran University of Engineering and
Technology, Jamshoro.
*** Assistant Professor, Department of Electrical Engineering, Mehran University of Engineering and Technology,
Jamshoro.

MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821] 167
TEXTURE RETRIEVAL WITH DESCRIPTORS BASED ON LOCAL FOURIER TRANSFORM: COMPARING THE RECTANGULAR AND
CIRCULAR NEIGHBOURHOODS

the texture description based on Markov random fields, 1.2 Circular Neighbourhood
local binary pattern operator, local Fourier transform, and
The absolute values of DFT coefficients are also
spatial grey level co-occurrence matrix, whereas second
susceptible to changes if the image is rotated, since the
category included texture description based on wavelets,
pixel values in the rectangular neighbourhood get altered
which use spatial-frequency representations. They [1]
in the rotated image. Therefore, the absolute values of the
found that those texture descriptors from the first category
perform better than those from the second category. They DFT coefficients are not fully rotation invariant.

also argued that the texture description based on pixel

Arof and Deravi [4] suggested a circular neighbourhood
neighbourhoods as small as 3x3 can prove even better
instead of the rectangular for extraction of similar texture
than those emphasizing the spatial frequencies, because
features based on 1-D DFT. It argues that because rotating
texture is the characteristic of the immediate pixel
the image alters the pixel values in the rectangular 9-pixel
neighbourhood. Ursani, A.A., et. al. [2] also found that
neighbourhood, and therefore, the features extracted from
texture descriptor based on local Fourier transform
the rotated and non-rotated image are dissimilar. On the
outperform the descriptor based on Gabor wavelets.
other hand, the pixel values in the corresponding 9-pixel
The texture descriptors based on discrete Fourier circular neighbourhood remain unaltered in the wake of
transform using rectangular neighbourhood and circular image rotation.
neighbourhood and were introduced in [3-4], respectively.
Since then, these texture features have been used for Fig. 2 shows the circular neighbourhood with the
several applications including segmentation of multibeam neighbours 1, 3, 5 and 7 (shown using empty circles) having
echosounder data [5]. This paper compares the texture non-integer pixel coordinates. Arof and Deravi [4]
descriptors extracted from the two neighbourhoods. interpolate these neighbours of non-integer coordinates
Results are presented on their overall retrieval accuracy in
using inverse Euclidean distance as explained in Equations
the form of precision vs. recall curve.
(1 and 2). For example, the pixel value at x1 (having the
1.1 Rectangular Neighbourhood coordinates X=0.707 and Y=0.707) is interpolated between
the four closest neighbours of integer coordinates, i.e.
DFT-based texture descriptor extracted from rectangular P0(1,0), P1(1,1), P2(0,-1), and C(0,0) in the rectangular
neighbourhoods was proposed in [3]. Fig. 1 shows the 8-
neighbourhood.
pixel sequence whose grey levels form an 8-value vector
to be transformed using the DFT. In an PxQ pixel image,
there are (P-2)x(Q-2) central pixels. The method proposed
in [3] computes DFT of the sequence of grey-levels of the
8 pixels around each central pixel in an image. Later, it
computes histograms of the magnitudes of the complex
values of X0, X1, X2, X3, and X4 DFT coefficients. These
histograms, called Local Fourier Histograms (LFH) were
used to describe the texture. The texture descriptor
consists of 40 values since each of the DFT coefficients is
quantized into 8 bins. Ursani, A.A., et. al. [6] introduced
phases of the DFT coefficients as rotation invariant texture
FIG. 1. THE 9-PIXEL RECTANGULAR NEIGHBOURHOOD IN
features in addition to the absolute values. THE SPACE DOMAIN

168 MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821]
 TEXTURE RETRIEVAL WITH DESCRIPTORS BASED ON LOCAL FOURIER TRANSFORM: COMPARING THE RECTANGULAR AND
CIRCULAR NEIGHBOURHOODS

4 Evaluating a texture descriptor requires two separate

1
xm   d m, n x n datasets; one for training the classifier, and other for testing
4 n1
 d m, n (1)
its performance. The following subsections explain
n1
separately the image database to be searched and the
Where m is a 2-D non-integer coordinate in the circular
query image set used the results reported in this paper.
neighbourhood, n is a 2-D integer coordinate in the
The retrieval experiments use the query images to train
rectangular neighbourhood, xm is the interpolated value
the classifier and then search the database of images for
at a non-integer pixel coordinate m, xn is value of one of
the closest matches.
the four closest neighbours in the rectangular
neighbourhood, dm,n is the inverse Euclidean distance 2.1 The Training and the Query Image-Set
between the coordinates m and n, given as in Equation
(2). We downloaded texture images of Brodatz album from a
webpage belonging to the University of Stavanger. The
1
d m, n  (2) downloaded images measured 640x640 pixels each.
mn
Realising that the Brodatz album in fact contains several
images of a given kind (woven wire, paper, canvas, cloth,
In addition, Arof and Deravi [4] suggest using mean and
stone, brick wall, water, flower, skin, etc) with varying
variance of the feature images of the DFT coefficients X0
zoom and lighting conditions etc., only a single image
through X4 instead of computing their histograms and
from each kind is included in the query image set, resulting
using bins values as features as in [3]. However, since this
in an image-set comprising 32 texture classes shown in
paper aims at comparing only the rectangular and the
Fig. 3.
circular neighbourhoods, we extract exactly the same
descriptor as the one explained in Section 1.1 from the two Each of these images is a texture class. The features vectors
types of neighbourhood. extracted from these 32 images are used for training the
NN (Nearest Neighbour) classifier to produce the results
The rest of the paper comprises Section 2 that explains the
presented in this paper. NN is a highly flexible classification
datasets used and the experimental setup, Section 3
process that does not involve any parametric modelling
compares the computational cost of the methods involved,
of the training data. This can offer both space and speed
Section 4 presents the results, and Section 5 concludes
advantages in very large problems [7]. Size of the original
the findings.
downloaded images was 6402 pixels each. These were
2. MATERIALS AND METHOD divided into 9 images of 2102 pixels each. One feature
vector was extracted from each of the 9 sub-images from
each of the training images. The mean of the 9 feature
vectors was used as the 10th feature vector representing
a class. This forms 32 LUTs (Lookup Tables), each
representing a texture class and was used as a query in
the process of retrieval. Each LUT contains 10 feature
vectors (i.e. representatives) of each texture class. Two
separate LUTs are formed for each training image, one
from each i.e. rectangular neighbourhood and the circular
FIG. 2. THE CIRCULAR NEIGHBOURHOOD AS SUGGESTED
IN [4] neighbourhood.

MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821] 169
TEXTURE RETRIEVAL WITH DESCRIPTORS BASED ON LOCAL FOURIER TRANSFORM: COMPARING THE RECTANGULAR AND
CIRCULAR NEIGHBOURHOODS

The image database to be searched is formed by dividing R

each of the 32 Brodatz images into 16 sub-images each recall  100 * (3)
T
measuring 1602 pixels. This gives a total of 16x32, i.e. 512
where R is the number of relevant images retrieved and T
sub-images. Each of these 512 images is then rotated to 0,
is the total number of relevant images in the database.
15, 30, 45, 60, 75, and 90° using bilinear interpolation and
Equation (4) gives the precision.
the central part from each measuring 1022 pixels is cropped.
This resulted in a database comprising 3584 images, R
precision  100 * (4)
containing 112 siblings from each of the 32 classes and N
oriented at seven different angles. Each of these database
where N is the total number of images retrieved.
images is represented by a single feature vector.

2.3 Image Retrieval The similarity criterion used for finding the matches is the
cross correlation coefficient. Since each query image is
Retrieval is the process of sorting the objects in the order represented by a set of 10 feature vectors, the coefficient
of relevance. In the image retrieval, an image database is of correlation is computed between each of the 3584 feature
searched for the N closest matches of a query image. vectors representing the database images and each of the
Performance of the retrieval process is measured in terms 10 feature vectors representing a query image. Hence 10
of precision and recall [8-9] computed as in Equations (3 similarity values are computed between each database
and 4). image and the query image.

FIG. 3. THE 32 IMAGES FROM BRODATZ ALBUM SELECTED TO FORM THE IMAGE DATABASE

170 MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821]
 TEXTURE RETRIEVAL WITH DESCRIPTORS BASED ON LOCAL FOURIER TRANSFORM: COMPARING THE RECTANGULAR AND
CIRCULAR NEIGHBOURHOODS

Retrieval is performed using the NN classification. The based features from circular neighbourhoods is thrice as
highest correlation coefficient is used for sorting the expensive as extraction of DFT-based features from
relevant images from the database. rectangular neighbourhoods.

3. COMPUTATIONAL COST ANALYSIS 4. RESULTS

Since the method of extraction from the two types of The results of image retrieval performed on the dataset
neighbourhoods basically remains the same, the only containing 3584 images with each of the 32 training images
difference is the additional cost of interpolation that is used as a query image. N closest matches of a query image
required in the use of circular neighbourhood. The cost of were sought from the dataset containing a total of 3584
computing 8-point FFT (Fast Fourier Transform) is given images, out of which only 112 were relevant to each query
by: and the rest were irrelevant. In the reported experiments,

C FFT  O N log 2 N  (5) N takes integer values from 220212 separated by 10. The
Since N=8, in our application, the complexity becomes: resulting 22 pairs of precision-recall values are plotted in

 
C FFT  O 8 log 2 8  O 24  (6)
the form of PR (Precision vs. Recall) curve.

which shows that one requires 24 multiplications and 24 Fig. 4 presents the results of the retrieval using the NN
addition operations to compute DFT of 8-pixel values classification. This PR curve shows the retrieval results
around a central pixel. averaged over all the 32 queries. It can be seen that the
LFH features extracted from the rectangular
Interpolating pixel values at a non-integer coordinate
neighbourhood perform better than those extracted from
requires 4 multiplications (i.e. M=4), 8 additions (i.e. A=8),
the circular neighbourhood. The curve representing
one division (i.e. D=1), and a Euclidean distance
rectangular neighbourhood shows greater recall for a given
calculation. One Euclidean distance calculations requires
precision and a higher precision for a given recall.
two additions (A=2), two multiplications (M=2), and a
square root operation (S=1). This gives the total of 6 5. CONCLUSION
multiplications (M=6), 10 additions (A=10), one division,
and one square root operation. Although the square root This paper presented the results of image retrieval on a
and division operations are more complex than the large image database containing 3584 images from 32
multiplication, but if one approximates each square root classes. Despite being computationally much more
and each division by two multiplications, we have M=10 expensive, the features extracted from interpolated circular
and A=10. Since one requires interpolating pixel values neighbourhood prove less efficient than those extracted
at 4 non-integer coordinates for computing DFT around from rectangular neighbourhood. The reason of this failure
each central pixel. Therefore, the complexity of four is that there is a pitfall in the hypothesis of the circular
interpolation operations around a single central pixel is neighbourhood itself. The interpolated neighbours also
given by: come from the same pixel values of the 9-pixel rectangular
neighbourhood that is unstable to the rotation. This paper
CINTP = 0(40) (7)
however used only amplitudes of the DFT coefficients.
This shows that interpolating the pixel values at non- Future work may investigate the performance of the phases
integer coordinates is twice as expensive as extracting the of the DFT coefficients extracted from rectangular and the
DFT-based features itself. Therefore, extraction of DFT- circular neighbourhoods as the texture features.

MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821] 171
TEXTURE RETRIEVAL WITH DESCRIPTORS BASED ON LOCAL FOURIER TRANSFORM: COMPARING THE RECTANGULAR AND
CIRCULAR NEIGHBOURHOODS

FIG. 4. RETRIEVAL VS. RECALL CURVES

ACKNOWLEDGEMENT [5] Intelmann, S.S., Cutter, G.R., and Beaudoin, J.D.,

"Automated, Objective Texture Segmentation of
Multibeam Echosounder Data-Seafloor Survey and
The authors are thankful to Mr. Trygve Randen, University
Substrate Maps from James Island to Ozette Lake,
of Stavanger, for providing the Brodatz Images for this Washington Outer Coast", Marine Sanctuaries
work. The images were downloaded from his web page: Conservation Series MSD-07-05, US Department of
http://www.ux.uis.no/~tranden/ Commerce, National Oceanic & Atmospheric
Administration, MD.31, November, 2007.

REFERENCES [6] Ursani, A.A., Kpalma, K., and Ronsin, J., "Texture
Features Based on Local Fourier Histogram: Self-
[1] Varma, M., and Zisserman, A., "Texture Classification: Compensation Against Rotation", Journal of Electronic
Are Filter Banks Necessary?", Proceedings of Conference Imaging, Volume 17, No. 3, pp. 030503-1-030503-3,
on Computer Vision and Pattern Recognition, Volume 2008.
2, pp. 691-698, 2003.
[7] Hastie, T., and Tibshirani, R., "Discriminant Adaptive
[2] Ursani A.A., Kpalma, K., and Ronsin, J., "Improved Nearest Neighbor Classification", IEEE Transactions
Texture Features Based on Local Fourier Histogram", on Pattern Analysis And Machine Intelligence, Volume
Wireless Networks, Information Processing and Systems, 18, No. 6, June, 1996.
Volume 20, pp. 19-28, 2008.
[8] Deng, Y., and Yang, S., "An Image Retrieval Approach
Based on Annular Color Moments and Energy Moments
[3] Zhou F., Feng J.F., Shi, Q.Y., "Texture Feature Based on
Features", Proceedings of 2nd International Conference
Local Fourier Transform", IEEE Transactions on Image
on Future Computer and Communication, Wuhan, May
Processing, Volume 2, pp. 610-613, 2001.
21-24, 2010.

[4] Arof, H., and Deravi, F., "Circular Neighbourhood and [9] Kekre, H.B., and Mishra D., "Digital Image Search &
1-D DFT Features for Texture Classification and Retrieval Using FFT Sectors of Color Images",
Segmentation", IEE Proceedings on Vision, Signal International Journal on Computer Science and
Processing, Volume 145, pp. 167-172, 1998. Engineering, Volume 2, No. 2, pp. 368-372, 2010.

172 MEHRAN UNIVERSITY RESEARCH JOURNAL OF ENGINEERING & TECHNOLOGY, VOLUME 30, NO. 1, JANUARY, 2011 [ISSN 0254-7821]