Retinex theoretical study
Download
Report
Transcript Retinex theoretical study
Approaches for Retinex and Their
Relations
Yu Du
March 14, 2002
Presentation Outline
Introductions to retinex
Approaches for retinex
The variational framework
Relation of these approaches
Conclusions
2
What Is Retinex
Lightness and retinex theory
E. H. Land 1971
Visual system of human
Retina: the sensory membrane lining the eye that receives the
image formed by the lens (Webster)
Reflectance and illumination
Edges and independent color senstion
3
Model of retinex (1)
S ( x, y ) R ( x, y ) L ( x, y )
The given image
The illumination part
The reflectance part
4
Model of retinex (2)
s ( x, y ) r ( x, y ) l ( x, y )
S
Input Image
Log
s
Estimate the
+
r̂
Exp
lˆ
Illumination
5
R̂
Three Types of Previous Approaches
Random walk algorithms
E. H. Land (1971)
Homomorphic filtering
E. H. Land (1986), D. J. Jobson (1997)
Solving Poisson equation
B. K. P. Horn (1974)
6
Random Walk Algorithms (1)
First retinex algorithm
A series of random paths
Starting pixel
Randomly select a neighbor pixel as next pixel on path
x1
Accumulator and counter
A( xi ) A( xi ) log( f ( xi )) log( f ( x1 ))
N ( xi ) N ( xi ) 1
7
Random Walk Algorithms (2)
Adequate number of random paths
Cover the whole image
Small variance
Length of paths
>200 for 10x10 image (D. H. Brainard)
8
Special Smoothness of Random Walk
The value in the accumulator
A( x)
log( f ( x)) log( f ( x ))
i
paths that
passed
pixel x
The illumination part
l ( x) log( G( x))
G( x) N f ( x1 ) f ( xN )
9
Homomorphic Filtering
Assume illumination part to be smooth
Apply low pass filter
H (u, v) ( H L )(1 e
D 2 ( u ,v )
c
D02
) L
10
Poisson Equation Solution (1)
Derivative of illumination part close to zero
Reflectance part to be piece-wise constant
Get the illumination part
Take the derivative of the image
Clip out the high derivative peaks
11
Poisson Equation Solution (2)
s
s T
(s)
0 other wise
Solve Poisson equation
lˆ (s)
Iterative method
Apply low-pass filter (invert Laplacian operator)
12
Comments on Above Approaches
Random walk algorithm
Too slow
Homomorphic filtering
Low-pass filtering first or log first?
More work needed to be done on Poisson equation
solving
13
Variational Framework
Presented by R. Kimmel etc.
From assumptions to penalty function
From penalty function to algorithm
14
Assumptions On Illumination Image
Spatial smoothness of illumination
Reflectance is not pure white
Illumination close to intensity image
Spatial smoothness of reflectance
Continues smoothly beyond boundaries
15
Penalty Function and Restrictions
Goal to minimize:
2
F (l ) l (l s) (l s) )dxdy
2
2
Subject to:
And
ls
l , n 0
on
16
Solve the Penalty Function (1)
Euler-Lagrange equations
F (l )
0 l (l s ) (l s )
l
And
ls
17
Solve the Penalty Function (2)
Projected normalized steepest descent (PNSD)
Iteratively to get illumination part
l j min{ l j 1 NSD G, s}
G l j 1 ( )(l j 1 s)
NSD
( G
G
2
2
(1 ) G )
2
18
Multi-resolution
Make PNSD algorithm converges faster
Illumination part is smooth
Coarse resolution image first
Upscale coarse illumination as initial of finer resolution
layer
Not multi-scale technique
19
Relationship of Different Approaches (1)
Random walk and Homomorphic filtering
R. Kimmel’s words on Homomorphic filtering
0
and remove constraint
ls
20
Relationship of Different Approaches (2)
Apply appropriate scaling on images,
Homomorphic filtering satisfies constrain
ls
and
0
Poisson equation approach:
( x, y ) (s)
21
Conclusions
Retinex is trying to simulate human vision process
Different approaches are from same assumptions
Implementation details are important for results
22
Thank You
March 14, 2002