multiscale retinex technique with color correction based on the retinex theory

• Edward Land’86

• There exists a discrepancy between the human vision system and the recorded color images.

• Dynamic range difference results in the loss of the essentials features from the recorded images.

• Improved fidelity of color images to human observation can be obtained by

(a) Computation that combines dynamic range

compression, color constancy and color rendition

(b) Color restoration.

Log

*

Gain/offset

CR

*

*

Log

Log Σ

I(x,y)

F1(x,y)

F2(x,y)

F3(x,y)

Σ

W1

W2

α

W3

+

+

+

+

-

MSRCR

||||||||||||||||

MSR

CRF

• First proposed design of Surround function by E.Land’86 was inverse square spatial surround

F(x,y) = 1/ [1+(r2 + c2)]

• The surround function was later modified in Gaussian form

by Hurlbert’89

F(x,y) = exp(-r2 / c2 )

Where

r- √ x2 + y2 and

c- Surround Space Constant

The Single Scale retinex is given by

Ri (x,y)=log Ii (x,y) – log [F(x,y) * Ii (x,y) ]

Where

F(x,y) = K exp(-r2 / c2 )--- Surround Function c- Scalar value and selection of K is thatr- √ x2 + y2

∫∫ F(x,y) dx dy =1

The multi-scale retinex is represented by

Ri (x,y)= Σ Wn { log Ii (x,y) - log[ F(x,y) * Ii (x,y) ]}

Where

n -- Scaling Factor

Wn – Weights (1/3 for each color channel of RGB)

N

n=1

http://dragon.larc.nasa.gov/pub/papers/multsclrtx.pdf

Limitations of the MSR:

• The Selection of the value of ‘c’ in equ(1) is critical.• The DRC results in the violation of Gray world algorithm• The region of constant color bleaches out as a result of DRC.

Gray World Assumptions:

Gray World Assumption states is that, given an image with sufficient amount of color variations, the average value of the RED, GREEN, and BLUE components of the image should average out to a common gray value.

The color restoration is calculated using the expression

Ci(x,y) = β{log[α Ii (x,y)] – log[ Σ Ii(x,y)]}Where

β- Gain Constantα- Controls the strength of non-linearity

The Final representation of MSRCR is represented as

R MSRCRi (x, y) = G [Ci (x, y) * RMSRi (x, y) + b]

Where G- Gain Constant and

b- Gain Offset value

s

i=1

Wn - 1/3 - Weight used in Multiscale Retinex

N – Number of Scale =3

C1, C2, C3 - Surround Constant – 15, 80,250 respectively

G - Final Gain – 192

b - Offset Value – 30

α – Strength of non-linearity – 125

β – Control gain constant - 46

Gaussian Surround Function

F(r)

Image Co-ordinate

Input

Output

Space Constant c=80

Inputs

Outputs

Space Constant c=80

Inputs

Outputs

c=15 c=80 c=215

Output of MSRCR:

Inputs

MSROutput

MSRCROutput

Tak at IRIS Laboratory:

Inputs

MSROutput

MSRCROutput

Tak at McGhee Tyson Airport:

Inputs

MSROutput

MSRCROutput

Few Examples:

Input Software My results

Tak at IRIS and McGhee Tyson Airport:

Input Software My results

The following image was presented as an example in the paper, the same

image is used as input to both the software available and my

implementation.

Test Image

http://dragon.larc.nasa.gov/pub/papers/multsclrtx.pdf

My Implementation

From Available Software