×

You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Title
Close-range imaging, long-range vision

A NEW ALGORITHM FOR IMAGE RETRIEVAL FROM IMAGE SEQUENCE
Hongchao Ma’, Deren Li
(Faculty of remote sensing, Wuhan University, China)
ABSTRACT
This paper describes the use of semivariogram as a parameter for image comparison which is a commonly used method in
content-based image retrieval. The authors first review various applications of spatial statistics to image and signal processing, and
recent literature of image comparison, with the emphasis to global image structure description and distance-based image retrieval
techniques. The difficulty arising in this field is the definition of image similarity. A new parameter based on semivariogram is putted
forward by the authors. Bearing in mind that semivariogram is a parameter not only describes the global structure of a data set but also
describes the local continuity of that data set, it is shown in the paper that semivariogram is suitable for global image comparison, and
can be used to reveal local features of the image as well. Based on this property, a new index for image similarity is constructed and a
practical program using it is developed. By applying the approach to a practical problem, the results show that this approach has the
following merits: (a) high sensitivity to structure differences of an image. (b) low computational complexity, and (c) high robustness to




lightening conditions.
Key Words: image comparison, semivariogram, spatial statistics, image similarity, image retrieval.
1. INTRODUCTION
Semivariogram is one of the basic concepts and tools from
spatial statistics, with the semivariogram estimation technique
and various kriging interpolation methods as its main contents.
The use of semivariogram to remotely sensed data analysis dates
back to late 1980's, when the pioneering works done by Curren
and Woodcock (Curren, P. J, 1988; Woodcock, C. E, 1988). The
earlier investigators mainly focused on exploring spatial data
pattern accompanied with remotely sensed data. From then on,
many new applications using spatial statistics for remotely
sensed data analysis have been carried out, including replacing
miss or bad data (Rossi, et al, 1994; Addink, et al, 1999),
signal-to-noise ratio assessment (Curren, P. J, 1989; Smith, G. M.,
et al, 1999), noise removal ( Green et al, 1988), classification
(Van der Meer, 1994; Carr, 1998; Atkinson, 2000; Chica-Olmo,
2000), anisotropic spatial modeling for remote sensing image
rectification (Cheng K.S. et al, 2000), scaling problems in
remotely sensed data (Myers, 1997; Collins, 1999), etc. A recent
review on spatial statistics application to remote sensing was
given by Stein, et al (1998).
The use of spatial statistics to close range photographic image
processing/signal processing is a more recent application.
Yfantis (1994) used a kriging-like iterative method with
changing interpolation neighborhoods and used the data
estimated in the previous iteration in order to estimate the data of
the present iteration. Their method was used for image
compression and it constituted a lossy compression algorithm.
The first step of their method is, as the usual step which must be
taken in order to use kriging, to estimate the semivariogram of an
image. Based on this they estimated the range of influence. If r is
the influence range estimated from the semivariogram of the

image, then the image was sampled using a square sampling
design with side length equals to a*r, where a is a factor less than
one. A kriging estimator was used to estimate the pixels not
included in the sample. The algorithm is lossy and the tests using
the image of Lenna shows that when high compression ratios are
desired, the algorithm produces recognizable images at a higher
degree than JPEG.
The applications of kriging to image sequence coding was used
by Decenciere, et al (1998). They first apply kriging to texture
coding and improve these preliminary results with an
optimization method that they have named “inverse kriging”.
Then they apply kriging and inverse kriging to motion vector
fields, which are by essence smooth within an object. In their
works, they combined kriging with morphological tools, which
to the author's best knowledge is the first one combining
morphological tools with kriging, two fields having close
relationship but never been used in image processing
cooperatively.
A recent application of kriging to signal processing was given by
Costa, et al (2000). A semi-parametric approach based on kriging
was suggested for nonlinear prediction It does not rely on any
specific model structure, which makes the approach much more
flexible than those based on parametric behavioral models.
Various examples were presented by them to illustrate the
robustness of the method and application on real data was
considered in the context of a noise-cancellation problem in
underwater acoustics.
This paper extends the use of semivariogram to image
comparison, which is a commonly used approach in
content-based image retrieval. The next section, section 2, will
briefly introduce the current state-of-the-art of image retrieval,
hongchao_ma(@263.net.cn or hemar@yahoo.com; phone 86-27-87664485; fax 86-27-87865402; Faculty of remote sensing, Wuhan University,

Wuhan, P.R. China. 430079
—504—
poi
can
ima
sec
par:
bet
sho
des
ins
Rec
ima
equ
info
Ber
cant
orge
retri
sinc
com
two
pers
The
197(
retri
text-
ima,
rese:
espe
vast
othe
cont
That
it d
impr
retri
In th
colle
appr
diffic
imag
visu:
inde
along
relate
(199
have
is the
Most
optio
by sk
navig
To se
imag
an im
this
simil
comp