Color Image Enhancement Using the Support Fuzzification in the

Framework of the Logarithmic Model

Vasile PATRASCU, Vasile BUZULOIU

Image Processing and Analysis Laboratory (LAPI),

Faculty of Electronics and Telecommunication

University POLITEHNICA of Bucuresti, Romania

Signal and Image Processing at LAPI

Research Seminar, Fermi Lab - November 2005
 Summary

The image enhancement method presented here uses point operations.

The particularity of the approach is that

- the logarithmic representation of images, i.e. the image values are

elements of another (Euclidean) space, not R (the real line) in the case of
gray images;
- and the image is structured using fuzzy partitions.

Simple though powerful methods for image enhancement can be obtained

using affine transforms, defined in the framework of the logarithmic model.
 Summary (cont)

Better results can be obtained if the image can be divided in statistically

uniform subimages i.e. defining a partition on the image support and allowing
a different transform in each subimage of the partition.

Using classical partitions we are faced with a block effect at the border of the
subimages. To avoid this drawback the classical partitions can be replaced by
fuzzy partitions. Their elements will be the “fuzzy windows”.

In each of them an affine transform will be defined using the fuzzy mean and
fuzzy variance computed for the pixels of the analyzed window.

The final image is obtained by summing up the images of every fuzzy window
in a weight way. The weights used are membership degrees, which define the
fuzzy partition.
 The Vector Space of Gray Levels

- the space of gray levels is the set: E = ( − 1,1)

we shall use the following notations:

- the addition: 〈 + 〉
- the scalar multiplication: 〈×〉

- the scalar product: ( |⋅⋅ ) E

- the norm: || . || E
 The Addition

The sum of two gray levels v〈1 v2 + 〉

v1 + v2
∀v1 , v 2 ∈ E , v1 〈 v2 += 〉 (1.1)
1 + v1v2

The neutral element: θ=0

The opposite of v ∈ E : w= −v

The Subtraction
 v1 − v2
∀v1 , v 2 ∈ E , v1 〈 v2 −= 〉 (1.2)
1 − v1v2

The Multiplication by a Scalar

The multiplication 〈×〉 of a gray level v by a real scalar λ:

∀v ∈ E , ∀λ ∈ R ,
( 1 + v ) λ − (1 − v ) λ
λ〈× 〉v =
(1 + v ) λ + (1 − v ) λ (1.3)

The addition 〈 +
and the scalar multiplication 〉 〈×〉 induce on E a real vector
space structure.
 The Fundamental Isomorphism

The vector space of gray levels ( E, 〈 , 〈 ) +× 〉

is isomorphic to the space of real

numbers ( R,+ ,⋅ ) by the function:

ϕ: E → R : ∀v ∈ E ,
1 1 + v 
ϕ( v ) = ln  
2 1 − v  (1.4)

The isomorphism ϕverifies:

∀v1 , v 2 ∈ E ,
 ϕ ( v1〈 v2) ϕ= ( v1) ϕ+ ( v2) + (1.5) 〉
∀λ ∈ R, ∀v ∈ E ,
ϕ( λ 〈× 〉v ) = λ ⋅ ϕ( v ) (1.6)

The Euclidean Space of Gray Levels.

The scalar product of two gray levels:

( |⋅⋅ ) E : E × E → R , ∀v1 , v2 ∈ E ,

( v1 | v2 ) E = ϕ ( v1 ) ⋅ ϕ ( v2 ) (1.7)
With the scalar product ( |⋅⋅ ) E the gray level space becomes an Euclidean space.

The norm
|| ⋅ || : E → R , ∀v ∈ E ,
E
+

|| v || E = ( v | v ) E =| ϕ ( v ) | (1.8)

For each image f

heaving the support D, the mean µϕ( f )
and the variance 2
σϕ (f)

are defined:

 1 
µ ϕ ( f ) = 〈  〈 f (+x, y) × 〉 〉
(1.9)
 c a(D) r d 
(x, y)∈ D
 2
| | yxf ),( µ〈
ϕ f )( | E | − 〉
σ ϕ f )( = ∑
2 (1.10)

yx ),( ∈ D
c ( D )
a r d

where card (D) is the cardinality of D.

The Logarithmic Model for the Color Space

Consider the cube E 3 as the color space. Let be q = ( r , g , b ) ∈E 3 , a color having the
components r (red), g (green) and b (blue).
The Addition
 ,
∀q1, q2 ∈ E 3

〈 qq = r121 2g1 g2b〈〈〈 21 ),,( (2.1)

The Subtraction
∀q , q ∈ E ,
1 2
3

〈 qq = r121 2g1 g2b〈〈〈 21 ),,( (2.2)

The Scalar Multiplication

∀λ∈R, ∀q ∈E , 3

λ〈×〉q = (λ〈×〉r , λ〈×〉 g , λ〈×〉b) (2.3)

The Euclidean Space of the Colors

The scalar product:

( |⋅⋅ ) E 3 3
3 : E × E → R, ∀q1 = (r1, g1, b1 ) , q2 = (r2 , g 2 , b2 ) ∈ E 3,

( q1 | q2 ) E 3 = ϕ (r1)ϕ (r2 ) + ϕ ( g1)ϕ ( g 2 ) + ϕ (b1 )(b2 ) (2.4)

The norm:

|| ⋅ || E 3 : E 3 → R + , ∀q =(r, g , b) ∈E , 3

|| q ||E 3 = ϕ 2 (r ) + ϕ 2 ( g ) + ϕ 2 (b) (2.5)

The Fuzzification of the Image Support

A gray level image is described by its intensity function:

f :D→E
(3.1)

where D ⊂ R 2 is the image support.

Without loss of generality, the rectangle
D = [ x0 , x1 ] × [ y0 , y1 ] (3.2)
can be considered as the image support.
The coordinates of a pixel within the support D will be noted ( x, y) .
Let there be

{ }
P = Wij | (i, j ) ∈ [0, m] × [0, n]
(3.3)

a fuzzy partition of the support D .

Consider the polynomials,
qxi : [ x 0 , x1 ] → [0,1] ,
 ( x − x 0 ) i ( x1 − x ) m−i
i
qxi ( x ) = C m
( x1 − x 0 ) m
(3.4)
qy j : [ y 0 , y1 ] → [0,1]
,
( y − y 0 ) j ( y1 − y ) n − j
j
qy j ( y ) = C n
( y1 − y 0 ) n
(3.5)
pij : D →[0,1]
,
pij ( x, y ) = qx i ( x) ⋅ qy j ( y )
(3.6)
m! n!
where i
Cm =
i!( m − i )! , Cnj =
j!( n − j )! and (i, j ) ∈[0, m] × [0, n] .

The membership degrees of a point ( x, y ) ∈ D to the fuzzy window W are given

ij

by the functions:

wij : D →[0,1]
,
( pij ( x, y ) ) γ
wij ( x, y ) =
n m
∑ ∑ ( pij ( x, y ) ) γ (3.7)
j =0 i =0
The membership degrees w ( x, y) describe the position of the point ( x, y) within
ij

the support D .
The parameter γ ∈(0, ∞) has the role of a tuning parameter offering a greater
flexibility in building the fuzzy partition P . In other words, γ controls the
fuzzification-defuzzification degree of the partition P.
Fig.1 The graphics of membership degrees for m=2, n=2 and λ=1
Fig.2 The graphics of membership degrees for m=2, n=2 and λ=2
Fig.3 The graphics of membership degrees for m=2, n=2 and λ=4
Fig.4 The graphics of membership degrees for m=2, n=2 and λ=16
For each window W the fuzzy cardinality card (W ) , the fuzzy mean ij ij µϕ( f , Wij )
and
the fuzzy variance σ ( f ,W ) of the image f are defined: 2
ϕ ij

card (Wij ) = ∑
( x , y )∈D
wij ( x, y )
(3.8)

 wi (jx, y) 
µ ϕ ( f ,Wi )j= 〈  〈 f (+x, y) × 〉 〉 (3.9)
 c (aWi )jr d 
(x, y)∈ D  
2
wi yx j),( | yxf ),(| µ〈 ϕ f,( Wi )|Ej | − 〉
2
σ ϕ f,( Wi )= j ∑ c (W )
a r d
(3.10)

yx ),( ∈ D i j
 The Enhancement Method for Gray Level Image
Let us consider these affine transforms on the images set F ( D, E ) , defined as
following: ψ : F ( D, E ) → F ( D, E ) , ∀f ∈F ( D, E )

ψ ( f ) λ= 〈 ( f 〈 τ ) × + 〉 〉
(4.1)

where λ ∈ R , λ ≠ 0 and τ ∈ E .

An image can be processed in two steps:

- a translation
f〈 τ + 〉 (4.2)

with a constant value τ, which leads to a change in the image brightness

- a scalar multiplication
 λ〈×〉 f (4.3)
by the factor λ - leading to a change in the image contrast.

The determination of the two parameters (λ, τ) will be made, so that the new
1
image will have the mean zero and the variance 3 .
σu
λ=
σϕ ( f ) (4.4)
1
where σu2 =
3 .

τ = 〈 µ ϕ (f ) − 〉 (4.5)

From statistical point of view, this means that the resulted image will be very
close to an image with a uniform distribution of the gray levels.
Fig. 5. The original image “news”, the affine transform and the enhanced image.

Fig. 6. The original image “cells”, the affine transform and the enhanced image
Fig. 7 The original image “lax”, the affine transform and the enhanced image.

Fig. 8 The original image “miss”, the affine transform and the enhanced image.
The fuzzy window W will supply a couple of parameters
ij (λ, τ) , which reflects
the gray level statistics:

σu
λij =
σϕ ( f , Wij ) (4.6)
where σu2 = 1 / 3 .

τ i =j 〈 µ ϕ ( f ,Wi )j− 〉 (4.7)

The function for the fuzzy window W : ij

 σu
ψ i (jf ) = 〈 ( f 〈 µ ϕ ( ×f ,Wi )− j) 〉 〉 (4.8)
σ ϕ ( f ,Wi )j

The transform ψenh is built as a sum of the affine transforms ψij

weighted with
the degrees of membership w : ij

n m
ψenh ( f ) = ∑ ∑wij 〈×〉 ψij ( f )
j =0 i =0
(4.9)
 Fig.9 The original image “med” and the enhancement with fuzzy partition.

Fig.10 The enhancement with classical partition and the enhancement without partition
Fig.11 The original image “street” and the enhancement with fuzzy partition.
 The Enhancement Method for Color Images

A color image is defined by three scalar functions red, green and blue:

r:D→ E (5.1)

g:D→E
(5.2)

b:D → E (5.3)

The image luminosity: l : D → E ,

1
l = 〈 ( r 〈 g 〈 b) ×++ 〉 〉 (5.4)
〉
3
The enhanced image renh , g enh , benh :

n m wσ
∑∑ ( )
i uj
re = n h 〈 r µ〈 ϕ (l,Wi ) j ×− 〉
(5.5) 〉
σj= 0 i= 0 ϕ (l,Wi ) j

n m wσ
∑∑ ( )
i uj
ge = n h 〈 g µ〈 ϕ (l,Wi ) j ×− 〉
(5.6) 〉
σj= 0 i= 0 ϕ (l,Wi ) j
 n m wσ
∑∑ ( )
i uj
be = n h 〈 b µ〈 ϕ (l,Wi ) j ×− 〉
(5.7) 〉
σj= 0 i= 0 ϕ (l,Wi ) j

Fig. 12 The original image “aerial1” and the enhanced with fuzzy partition.
Fig. 13 Enhanced without partition and the enhanced with classical partition.
Fig. 14. The original image “island” and the enhanced with fuzzy partition.
Fig. 15. The original image “aerial2” and the enhanced with fuzzy partition.
Fig. 16 The original image “player” and the enhanced with fuzzy partition.
Fig. 17 The original image “Egyptian bath” and the enhanced with fuzzy partition.
Fig. 18 The original image “puerta”, the enhanced with fuzzy partition and enhanced without
partition.

REFERENCES
[1] V. Patrascu, V. Buzuloiu, “Color Image Enhancement in the Framework of
Logarithmic Models”, The 8 th IEEE International Conference on Telecommunications,
P P

Vol. 1, pp. 199-204, ICT2001, June 4 - 7, 2001, Bucharest, Romania.

[2] V. Patrascu, V. Buzuloiu, “A Mathematical Model for Logarithmic Image Processing”,
The 5th World Multi-Conference on Systemics, Cybernetics and Informatics, Vol 13, pp.
P P

117-122, SCI2001, July 22-25, 2001, Orlando, USA.

[3] V. Patrascu, V. Buzuloiu, The Affine Transforms for Image Enhancement in the Context
of Logarithmic Models, Proceedings of International Conference on Computer Vision and
Graphics, Vol. 2, pp. 596-601, ICCVG 2002, September 25-29, 2002, Zakopane, Poland.
[4] V. Patrascu, Logarithmic Image Enhancement Using the Image Support Fuzzification,
Symbolic and Numeric Algorithms for Scientific Computing, MIRTON Press, Timisoara,
2002, Proceeding for 4 th International Workshop SYNASC 02, pp 253-262, October 9-12,
P P

2002, Timisoara, Romania.

[5] V. Patrascu, V. Buzuloiu, Modelling of Histogram Equalisation with Logarithmic Affine
Transforms, Recent Trends in Multimedia Information Processing, World Scientific
Press, 2002, Proceedings of the 9 th International Workshop on Systems, Signals and
P P

Image Processing, IWSSIP’02, pp. 312-316, November 7-8, 2002, Manchester, United
Kingdom.
[6] V. Patrascu, V. Buzuloiu, C. Vertan, Fuzzy Image Enhancement in the Framework of
Logarithmic Model, In Fuzzy Filters for Image Processing, pp 219-237, Springer Verlag,
Germany, March, 2003 .
[7] V. Patrascu, Gray Level Image Enhamcement using the Support Fuzzification in the
Framework of Logarithmic Models, In proceedings of the International Conference
Model-based Imaging, Rendering, Image Analysis and Graphical Special Effects, Mirage
2003, pp. 123-128, INRIA Rocquencourt, 10-12 March, 2003, France.
[8] V. Patrascu, V. Buzuloiu, Color correction method using logarithmic operations,
Proceedings of the 7th International Conference on Engineering of Modern Electric
P P

Systems, EMES’03, Oradea, Romania, pp. 150-155, May 29-31, 2003.

[9] V. Patrascu, Color Image Enhancement Using the Support Fuzzification, In Fuzzy Sets
T T

and Systems – IFSA 2003, Vol LNAI 2715, Springer-Verlag Berlin Heidelberg 2003,
Proceedings of the 10th International Fuzzy Systems Association World Congress, pp.
412-419, June 30 – July 2, 2003, Istanbul, Turkey.
[10] V. Patrascu,V. Buzuloiu, Color Image Processing Using Logarithmic Operations, In
Proceedings of The IEEE International Symposium on Signals, Circuits and Systems,
SCS 2003, pp. 133-136, July 10-11, 2003, Iasi, Romania.
[11] V. Patrascu, V. Buzuloiu, The mean dynamic range optimization in the framework of
logarithmic models, Advanced Topics in Optoelectronics, Microelectronics, and
Nanotechnologies, Proc. SPIE, Vol. 5227, Oct. 2003, pp. 73-80.
[12] V. Patrascu, Image Enhancement Method Using a New Measure for Mean Dynamic
Range, Proceedings of the conference IPMU’04, Information Processing and
Management of Uncertainly in Knowledge-Based Systems, pp. 1921-1928, July 4-9,
2004, Perugia, Italy.
[13] V. Patrascu, Color Image Enhancement Using the Lrgb Coordinates in the Context of
Fuzzy Sets Theory, The 8th World Multiconference on Systemics, Cybernetics and
Informatics, Vol. 6, July 18 - 21, 2004 Orlando, Florida, USA
[14] V. Patrascu, Fuzzy Enhancement Method Using Logarithmic Models, Proceedings of the
13th IEEE International Conference on Fuzzy Systems, FUZZ-IEEE2004, Vol. 3, pp.
P P

1431-1436, Budapest, Hungary, 25-29 July, 2004.

[15] V. Patrascu, Color Image Enhancement Method Using Fuzzy Surfaces in the Framework
of the Logarithmic Models, In vol. Applied Computational Intelligence, World Scientific
Press, 2004, Proceedings of the 6th International FLINS Conference, FLINS 2004, pp.
403-408, Blankenberge, Belgium, 1 - 3 September, 2004.