03-05-2013, 02:47 PM
An Unsupervised Segmentation Framework For Texture Image Queries
An Unsupervised Segmentation.pdf (Size: 1.14 MB / Downloads: 25)
Abstract
In this paper, a novel unsupervised segmentation framework
for texture image queries is presented. The proposed
framework consists of an unsupervised segmentation
method for texture images, and a multi-filter query strategy.
By applying the unsupervised segmentation method on
each texture image, a set of texture feature parameters for
that texture image can be extracted automatically. Based
upon these parameters, an effective multi-filter query strategy
which allows the users to issue texture-based image
queries is developed. The test results of the proposed framework
on 318 texture images obtained from the MIT VisTex
and Brodatz database are presented to show its effectiveness.
Introduction
Segmentation is an important part of the computer vision
and image analysis, wherein regions of interest are identified
and extracted for future processing. The definition of
suitable similarity and homogeneity measures is a fundamental
task in many important applications, ranging from
remote sensing to similarity-based retrieval in large image
databases such as the query by image content (QBIC) system
[4].
Unsupervised Texture Segmentation
In the proposed unsupervised texture segmentation
framework, the partition and the class parameters are treated
as random variables. The method of partitioning a still image
starts with a random partition and employs an iterative
algorithm to estimate the partition and the class parameters
jointly [3].
Initial Partitions for Segmentation
The proposed segmentation method starts with a randomly
generated initial partition. Hence, different initial
partitions yield to different local minima. The smallest local
minimum among them gives the desired solution though
it may not be the global minimum. In the proposed framework,
a number of local minima (e.g., 20) are computed
and the smallest local minimum is used. Since the computational
requirement for each local minimum is very little,
the overall computation needed for the best local minimum
is not much. Two methods are used to generate those
twenty initial partition candidates. By the straight-line partition
method, the area of the original texture images is partitioned
by an arbitrarily generated straight-line across the
whole image area. Different areas separated by the straightline
represent different classes. In many cases, the randomly
generated straight-line partitions are good enough to get the
desired initial partition, but in many other cases it cannot
work well.
Query Strategy
After the segmentation on each texture image, a set of
parameters for each image is obtained automatically. Some
of these parameters are selected for query use. Since the
proposed segmentation method uses the functions of the
spatial coordinates of the pixels as the mathematical description
of a class, those parameters related to spatial information
should be able to represent the spatial distribution
features of textures.
r Parameter sut : After the segmentation, each pixel
within a texture has its class identification. For example,
the class identification for each pixel is either 1 or
2 when there are two classes. As mentioned earlier,
each class is parameterized by a vector of parameters
)85 6:9 "<;=;>;=" 56@? -A . In other words, this parameter vector
contains not only the spatial distribution information of
the texture, but also the information of intensity values
within that class. Furthermore, among the four parameters
in the vector, 576:9 is usually far more larger than
the other three. Therefore, given the number of classes
is 2, two svt parameters (one for each class) are obtained
for each texture.
Test Results and Discussions
Image Retrieval Results
In order to test the performance of the proposed framework,
318 natural texture images mostly obtained from the
MIT VisTex Texture database and Brodatz database are
used. For the images from Brodatz, we partition each of the
512 512images into 6 subimages (with overlap). Each
texture image is of size 240 rows and 180 columns. In the
proposed framework, the similarity query is used. An example
of the query looks like “Show me more texture images
which are similar in texture patterns with the query image.”
Conclusion and Future Work
In this paper, an unsupervised segmentation framework
for texture image queries was proposed. By using a novel
and effective segmentation method, a set of feature parameters
for each class within an image is extracted automatically
without any user interference. Based on these feature
parameters, the proposed framework supports texture image
queries effectively. Moreover, a multi-filter mechanism
is used in the query procedure to greatly reduce the number
of image candidates and at the same time, reduce the query
processing time. Furthermore, applying the segmentation
method on partitioning the natural image also gives good
results.
One of the potentials of the proposed segmentation
method is that it can also deal with the situation of multiple
classes (more than two). The idea is to consider the
number of classes | as another random variable. Our future
work will focus on generalizing the proposed framework to
handle the cases when the number of classes is more than
two so that it can partition the image more reasonably and
precisely, which is essential to the accuracy of the queries.