Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 82

and

a ∈ A w ( a )= 1

(3.26)

, ʸ )= a ∈ A P a ( s , ʸ ) w ( a )

P ( s (3.27)

where E a ( represent the energy and probability of the edge flow

computed from image attribute a

s , ʸ ) and P a ( s , ʸ )

∈ {intensity/color, texture, phase}. w ( a ) is the

weighting coefficient associated with image attribute a .
For a given color image, the intensity of the edge flow can be computed in
each of three color bands

( R , G , B ) and the texture edge flow can be calculated

from the intensity I 3 . Then the overall edge flow can be

obtained by combining them as in Eqs. ( 3.25 ) and ( 3.27 ) with A

=( R + G + B ) /

{red, green,
blue, texture}. Each location s in the image is associated with the three parameters:
{ [

=

E ( s , ʸ ) , P ( s , ʸ ) , P ( s , ʸ + ˀ )] | 0 ≤ ʸ < ˀ } . Given these parameters, Eq. ( 3.24 )

is utilized to firstly obtain the parameter . Then, the edge flow vector F is

identified by Eq. ( 3.23 ). The resulting F is a complex number with its magnitude
representing the resulting edge energy and angle representing the flow direction.
The basic idea of the Edge Flow method is to identify the direction of change
in the attribute discontinuities at each image location. The Edge Flow vectors
propagate from pixel to pixel along the directions being predicted. Once the
propagation process reaches its stable state, the image boundaries can be detected
by identifying the locations which have non-zero edge flows pointing to each other.
Finally, boundary connections and region merging operations are applied to create
closed loop regions and to merge these into a small number of regions according to
their color and texture characteristics.

ʘ ( s )

3.4.3 Knowledge-Based Automatic Region of Interest

The definition of ROI is highly dependent on user needs and perception. However,
specific to the current application for photographic collections, a photographer
usually creates a photograph with a single focus point at the center of the picture.
Based on this assumption, we can effectively attain ROI by associating it with
the objects located at the center of photographs. Let

S = {R i , i = 1 ,..., N |R i ∩

R j = ∅ , i = j } be a set of regions generated by the Edge Flow model from one

image, where R i is the i -th region and N is the number of regions. Let W m × n be a

predefined rectangular window of size m × n pixels, whose center is located at the

center of the input image. Also, let be a set of label for regions that are located

either partly or completely inside the W m × n window, e.g.,

W

W = { i | R i ∩ W m × n = ∅ } .

ROI is defined as a collection of regions which are members of W :

S = R i (3.28)

i ∈W