1240 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 9, NO. 7, JULY 2000
Fig. 2. (a) Gaussian pyramid representation of a synthetic image with
SNR = 10
dB, from layer 3 (top image of
8
3
8
size) to the bottom layer (bottom image of
256
3
256
size). Note that for display purposes the images of the upper layers were enlarged. (b) Pyramidal segmentation of the image of Fig. 2(a) from layer 4 into
the bottom layer. Different gray values of the image elements correspond to different labels. (c) The fuzzy image of the class disk from layer 4 to the bottom layer.
The fuzzy image of the class background is not displayed. (d) Defuzzification of the fuzzy images of (c). Each pixel of a layer in an image is assigned to one class.
combination of convolution and sub-sampling a pyramid repre-
sentation of an image can be obtained. For an image of original
size
, the pyramid representation of an image at the
th layer of the pyramid can be described as
(1)
where
is the gray value of an image at the posi-
tion in the
th layer of thepyramid, is a filter coefficient,
and are the filter sizes. In a Gaussian (low-pass) pyramid
[16], the pyramid is constructed by repeated application of local
averaging and sub-sampling; see Fig. 2(a). The pyramid can also
be built by the difference of low-pass [17] or Laplacian operator,
by which specific structures in an image such as blobs (maxima)
and ridges can be detected. The pyramid may contain an image
series in each layer rather than a single image. For example, a
coarse representation of an X-ray or MR image series, acquired
over a cardiac cycle, can be represented at different layers of the
pyramid, each of which containing all images of the series.
2) Pyramidal Segmentation: Like other segmentation tech-
niques, the primary goal of pyramidal segmentation is to find
the location and extent of all objects that exist in an image. A
number of pyramidal segmentation techniques exist. Generally,
the blurred images represented in each layer of the pyramid are
used for segmentation. Segmentation is applied in two stages.
Firstly, a parent-child spatial relationship between the image el-
ements of two adjacent layers is defined. Secondly, this rela-
tionship is evaluated by means of a similarity measure. Since
different segmentation results can be obtained by different def-
initions of the spatial relationship and the similarity measure,
these two concepts are described in more detail in the following:
Spatial relationship between the image elements of two suc-
cessive layers of the pyramid describes the family relationship
between these elements. For example: in a quadtree represen-
tation each image element in the layer
is the parent of
the four nearest image elements (children) in layer
, so that
each child has only one parent and therefore no ambiguity ex-
ists within the parent-child relationship. In a linked pyramid this
relationship is ambiguous. The children of a layer can belong to
different parents in the upper layer.
Similarity between a child image element andits possible par-
ents describes how similar they are. By using features of image
Authorized licensed use limited to: YEUNGNAM UNIVERSITY. Downloaded on February 4, 2009 at 00:16 from IEEE Xplore. Restrictions apply.