In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
348
HIGH RESOLUTION IMAGERY RETRIEVAL
ON THE BASIS OF SKETCH-MODELLING
N. M. Kovalevskaya 3 , K.A.Boenko 3
3 Institute for Water and Environmental Problems SB RAS, Barnaul, 656038, Russia - knm@iwep.asu.ru
KEY WORDS: Environment, Monitoring, Retrieval, Global-Environmental-Databases, Model, Content-based, High resolution
ABSTRACT:
Recent technological advances have made it possible to process and store large amounts of image data. The most impressive
example is the accumulation of image data in scientific applications such as satellite imagery. However, in order to realize their full
potential, tools for efficient extraction of information and for intelligent search in image data bases need to be developed. The paper
describes a new approach to image data retrieval that allows queries to be composed of textured patterns. The textured pattern is
converted into a feature representation of reduced dimensionality which can be used for searching similar-looking patterns in the
database. This representation is obtained by the texture sketch model based on Gibbs random field approach for high resolution
satellite imagery. Experimental results are presented, which illustrate that the proposed representation preserves the perceptual
similarities, and provides an effective tool for content-based satellite image retrieval. As well visual and manual image-
interpretations produce similar outlines of geographical units.
1. INTRODUCTION
Content-based image retrieval has been a topic for research in
the last decades. A number of overviews on image database
systems and image retrieval have been published, see e.g.
(Veltkamp, 2001; Braveen, 2009). Despite of ongoing research
and numerous studies, no effective features have as yet been
generally accepted for image retrieval from currently available
satellite image data base (SIDB), especially high resolution
SIDB (HRSIDB). Quite a few investigations on satellite image
retrieval systems are focused on either retrieval by keywords
(Smith, 1996) or discerning of very specific features (Kelly,
1995; Wang, 2001). To query an image in accords to user-
defined pattern, a pattern's attribute vector that typically relates
to 3 descriptive characteristics (i.e. color, shape and texture) is
calculated. The next step implies similarity pattern retrieval.
Different Geographical Units to be identified or Mapped
(GUMs) differ by their contents in SIDB . Depending upon the
specific aim of an interpretation, a GUM may be for example a
vegetation patch, a part of the sea surface with a uniform wave
pattern, a patch of homogeneous land-use and so on.
Color characteristics (histograms) don't allow for spatial
delimitations ; therefore, color histogram based image retrieval
of (HR)SIDB-images often leads to erroneous query results.
One can hardly obtain effective query results by the use of
shape characteristics, since shapes of natural GUMs are
extremely diverse and complicated. GUMs presented in SIDB
by more or less homogeneous patterns in grey level scale,
which are decisive for content-based image retrieval. Spatial
homogeneity of GUMs in high resolution imagery is directly
related to their textural features.
The presented research aims at
1. a formalised description of textural features, and
2. the development of a model presentation that
describes homogeneity of textural patterns in terms of
probabilistic self-similarity.
The results of content-based image retrieval using the proposed
model, are also presented in this paper.
2. SKETCH TEXTURE MODEL
Precise definition of texture doesn't exist yet that is evidence of
the term complexity. Texture (from Latin textura means
"weaving" or "structure") relates to the specific structure of
visual or tactile characteristics of individual objects (Gimel'farb,
1999). In a broad sense, texture defines the structure of an
object with respect to the pattern along which its components
are arranged. For human perception, texture is the specific,
spatially repeated (micro- and macro-) structure: the spatial
arrangement of major surface components.
For satellite imagery the texture is presented as spatial
interactions of raster elements and their spatial arrangement.
Visually, such spatial interactions are presented as repeated
changes of grey levels in a proximity window. The
Gibbs/Markov models of piecewise-constant regions of the
Earth surface are an effective representation of textural objects
(Kovalevskaya, 2002). In fact, these models are rather flexible;
they allow to "seize" essential parts of visual information
presented by piecewise-homogeneous images. The key
parameters of the Gibbs model for obtaining metadata of visual
pattern content are the following:
1. the size of the proximity window of neighbouring
pairwise elements,
2. the structure of neighbouring elements representing
major visual pattern content,
3. the significance of each element in the structure.
Let us suppose that natural textures in high resolution images
possess a spatial self-similarity that can be expressed as the
frequency of pairwise elements. Probabilistic self-similarity of a
homogeneous texture pattern means that all probable
combinations of signals in pairwise cliques are considered as
having different likelihood of occurrence on the textural pattern
(Gimel’farb, 1999). Then, one can state that two patterns are
