ALGORITHM FOR CONTENT DATABASES SEARCHING USING

IMAGE DECOMPOSITION WITH GABOR FILTERS

Muguras Mocofan
Department of Communications, Electronics and Telecommunications Faculty
Politehnica University of Timisoara

ABSTRACT: The Gabor filters offer good performance in the field of content-based image indexing and
retrieval. Here, I present an algorithm for indexing and retrieval, which use the similarities, properties of human
visual system for textured images. This system is very useful for clustering textured images using similarities
offered by a multichannel decomposition with Gabor filters. The area of applications is very wide: multimedia
documents, transaction systems, medical application, query and browsing.

Introduction

A major problem in dealing with large image databases is the efficiency of retrieval. One of
the key issues in achieving such efficiency is the design of a suitable indexing scheme.
Content-based image retrieval can be only effective if it entails reducing the search from a
large and unmanageable number of images to a few that the user can quickly browse. The idea
is to work with descriptions based on properties that are inherent in the images themselves.
The idea behind this is that the natural way to retrieve visual data is by query based on the
visual content of an image: the patterns, colors, textures, and shapes of image objects, and
related layout and location information. The motivation of this process is in close relations
with the solid human eye’s capacity of accomplishing discrimination of images based on
textures. The system performs a multichannel decomposition using different orientations of
the Gabor filters for a good discrimination.
The area of applications is very wide: multimedia documents, transaction systems, medical
applications, query and browsing.

Texture Gabor filtering

The method is based on a set of features, which operate in parallel during image
decomposition into a collection of sub images. Individual filters are designed in such a way
that these simultaneous actions should concentrate on a frequency area and on local spatial
interaction [1]. Because image texture is viewed differently according to analysis resolution,
segmentation techniques that use Gabor filters based on a single frequency are not entirely
satisfactory from the quality retrieval point of view [2]. That is the way for an indexing
method that uses a multichannel filtering technique being proposed, frequency space division
being made according to perception by human visual system [3]. It is very hard to go
throughout the entire frequency space and for different orientations from the calculation point
of view [4]. That is why it is proposed for decomposition only through few channels. The
system output is optimized from the point of view of average square error and takes into
consideration 25 Gabor filters outputs.
Algorithm description

The goal of the proposed algorithm is to find a similarity between images, and to clustering
the most closed images in concordance with the human visual system. For that reason are
used the image decompositions of 25 Gabor filters (these filters are a good model for the
human visual system) in the following way:
- Every feature is compared with all features and the most closed features, which respect an
imposed threshold, are noted with a flag.
- The clustered images are the images with the same flags.
- It is possible to increase the number of the retrieved images by a simple extension of the
primary algorithm, changing the restriction „images with the same flags” with „images
with the same number of flags”.
I use different techniques for image clustering. Mixing the trees techniques with the human
interpretation of features the speed of the searching process increases. There are some
important features for the human visual system. For example, the mean, and few orientations
of the Gabor filters are representative (the horizontal and vertical direction, the diagonals) [5].
If the restrictive condition for this feature is not satisfied, the search process is stopped and
other images features are compared. The computational volume will be reduced.
In many papers, I founded different methods witch propose metrics for discrimination. This
metrics cannot offer a good discrimination because only two different features with very big
response minimize the contributions of other features (with small response).
The same problem is with the similarity measure of histograms. The similarity is defined in
equation (1):

N  Hist Im 1 (c j ) − Hist Im 2 (c j ) ⋅ N 
∑j = 1  1 − 
1
Simcolor ( Im 1, Im 2) =
( )  (1.)
 max N ⋅ Hist Im 1 (c j ), N ⋅ Hist Im 2 (c j )
N


The histograms are global features for an image and it is possible to obtain a very good
similarity for two different images but with the same histograms. For example an image with
vertical black and white bars and an image with horizontal black and white bars are two
distinct images, but, the histograms can be the same.

Experimental results

In my simulations I use other two methods, Euclidian metrics and the similarity between
histograms so as to compare the results of the proposed algorithm. Figure 1 represents the test
image used for finding the most similar images with that.

Figure 1. The test image

Figure 2. The results of retrieval using Euclidian metrics

Using similarity measure between histograms, the results are much better. In figure 3 there are
the results of retrieval using similarity.

Figure 3. The results of retrieval using similarity between histograms

The results of the proposed algorithm, for different value of threshold are presented in figure
4, 5 and 6. There are presented only three values for threshold ( t1 = 1.5, t2 = 1.2, t3 =1.05)
because the conclusions are very clear presented by the results.

Figure 4. The results of the proposed algorithm for t1 = 1.5

Figure 5. The results of the proposed algorithm for t2 = 1.2

The algorithm was tested with 300 images from a database. The spectrum of this database was
very large: real images from nature, animals, peoples, textures, artificial images. After many
experiments some conclusions could be extracted:
- The image decompositions by Gabor filters is a similar approach with the visual human
interpretation system. It’s not necessary to use many filters. Using a large number of filters
the computational volume will grow, and the algorithm will be very slowly.
- An important step in the algorithm design is to make a very right decomposition with
Gabor filters, to cover the entire space [6].
- A large range for threshold is not representative for clustering images. Using an iterative
system, the threshold can be reduced step by step. The resulted images are more and more
similar. A very small threshold is not a solution, because the number of features used is
also too small and only few features don’t describe very well an image.
The similarity for the human visual interpretation system is collections of features which have
a little tolerance. The global interpretation of image, like similarity between histograms, is
useful, but it is not a sufficient method.

Conclusions

An index in a database consists in a collection of entries, one for each data item, containing
the value of a key attribute for that item and a reference pointer that allows immediate access
to the item. Selection, derivation, and computation of image features that provide more useful
query expressiveness are the main goal for this work. I don’t implement the methods that are
based on exact matching; I use an implementation of retrieval methods based on similarity to
obtain clustered images.

Bibliography

[1] Dunn D., Higgins W. – Optimal Gabor Filers for Texture Segmentations, IEEE Transactions
on Image Processing, vol 4, pp. 947-964 july 1995.
[2] Cocquerez J. P., Philipp S. – Analyse d’images: filtrage et segmentation, Mason, Paris, 1995
[3] Trygve R. – Filter and Filter Bank Design for Image Texture Recognition, PHDTesis 1997.
[4] Bovik, Clark M. – Multichannel texture analysis using localized spatial filters, IEEE Trans.
Patt. Anal. Machine Intell vol. 12, pp. 55-73, Jan. 1990.
[5] Mocofan M., Pescaru D. – A parallel implementation of a algorithm for content-based image
indexing and retrieval using texture featres, Proceedings of International Conference
“Communications 2000”, 7 - 9 december 2000, Bucharest, Romania, pp. 85-88, 2000.
[6] Mocofan M., Pescaru D. – A parallel implementation of the Gabor image segmentation
algorithm using textures, Proceedings of International Symposium on System Theory,
Automation, Robotics, Computers, Informatics, Electronics and Instrumentation, “SINTES-
10”X-th edition, 25-26 May 2000, Craiova, Romania, pp. A201-204, 2000.