Image Retrieval Based on Color-Spatial Distributing Feature - Multimedia and Signal Processing - page 83

For a given image I ,

Ixy denotes color value of the pixel of position (, )

{ }

(, ) xy ,

i Ax y

= (, )(, ) x y ∈ I I x y ,(, ) = i ,1 ≤ i n ≤ is the set of all pixels with

color i , n is the number of color category,

n expresses the number of pixels in the

A . Let

cxy = (, )

A ,

be the spatial central position of all the pixels in the

i i i

x and y are defined as

where

1 1

  (8)

x = x y = y

i i

n n

(,) xy A ∈ (,) xy A ∈

i i

i i

ρ

expresses distributing cohesion of the same or similar color and is defined as

i

n

1



ρ () i = ( ( x − x )

2

+ ( y − y ) )

2 − 1

(9)

k i k i

n

k = 1

i

Where (,
k

xy

A . If

ρ () i

is very big, that
means the color distribution is relatively scattered. On the contrary, it means the color
distribution is relatively concentrated.

is the coordinates of pixel in set

k

4 Similarity Measures for Ranking Images

I I

Given a image , the feature vector of is

' ' ' '

( hI ( ) ρ ( I ) ⋅⋅⋅ hI I hI I ( ), ρ ( ), ⋅⋅⋅ , ( ), ρ ( ), ) h ( I )

, in which is the

1 1 i i n n i

weighted frequency of color value i ,

ρ () I

expresses the cohesion of all pixels with
color value i , n is the number of color category. Assume that a query image

i

I

'

I , the distance d between the two image is changed as

d .

and a database image

hI hI

'

() −

'

( )

ρ () I − ρ ( ) I

n



1 i 2 i

' 1 i 2 i

dII (, ) = ×

(10)

h 12

' '

1( + hI hI ) + ( ) 1( + ρ I ) + ρ ( I )

i = 1

1 i 2 i 1 i 2 i

Similarity Measures of two images include two parts, the first term of equation (10)
takes into account weighted color histogram, which is based on the consideration
importance of pixel position, namely that pixels with larger color contrast are more
likely to attract human attention than ones with smaller color contrast. The second
term of equation (10) takes into account the similarity of the pixels with similar color
distributing cohesion. The smaller

'

h dII is , the more similar the images

(, )

I

12

I are. The image retrieval methods deliver the database images with the small-

and

'

h dII to the user.

(, )

est

12