Backlight Compensation Using Retinex Algorithm...2015/11/04 · ISL Lab Seminar INDEX 2019-04-10 01...
Transcript of Backlight Compensation Using Retinex Algorithm...2015/11/04 · ISL Lab Seminar INDEX 2019-04-10 01...
Backlight Compensation Using RetinexAlgorithm
Han-Sol Kang
ISL Lab Seminar
INDEX
2019-04-10
01 Retinex Theory
02 Retinex Algorithm
03 Experiment Results
04 Conclusion
2
2019-04-10
01 Retinex Theory
3
∑ What’s the Retinex Theory
* Land, Edwin H., and John McCann. "Lightness and retinex theory." JOSA 61.1 (1971): 1-11.
*
Retinex = Retina + Cortex
Both the eye and the brain are involved in the Image processing.
A B=
If we just use the eye,
A B≠
But actually we use both the eye and the brain
2019-04-10
02 Retinex Algorithm
4
∑ Retinex Algorithm
* Jobson, Daniel J., Zia-ur Rahman, and Glenn Woodell. "A multiscale retinex for bridging the gap between color images and the human observation of scenes." Image Processing, IEEE Transactions on 6.7 (1997): 965-976.
*
Weber-Fechner’s law Land의시각적모델
),(),(),( yxLyxRyxI iii +
2019-04-10
02 Retinex Algorithm
5
∑ Single-Scale Retinex (SSR)
Log
*),(F yx
),(I yxi
-
SSR∑
+
Log
)),(*),(log(),(logR yxIyxFyxI iii
* Jobson, Daniel J., Zia-ur Rahman, and Glenn Woodell. "A multiscale retinex for bridging the gap between color images and the human observation of scenes." Image Processing, IEEE Transactions on 6.7 (1997): 965-976.
outputRetinex:),(
functionSurround:),(
)3,2,1(colorth:),(
yxR
yxF
iiyxI
i
i
*
2019-04-10
02 Retinex Algorithm
6
∑ Surround Function
Log
*),(F yx
),(I yxi
-
SSR∑
+
Log
* Jobson, Daniel J., Zia-ur Rahman, and Glenn Woodell. "A multiscale retinex for bridging the gap between color images and the human observation of scenes." Image Processing, IEEE Transactions on 6.7 (1997): 965-976.
*
)(1
1),(
22 cryxF
2
2
),( c
r
eyxF
ConstantspaceSurround:
22
c
yxr
• Surround function by E.Land’86 (Inverse square spatial surround) • Surround function by Hurlbert’89(Gaussian)
2019-04-10
N
n
n
N
n
nnMSR WRWRii
11
1,
7
02 Retinex Algorithm
∑ Multi-Scale Retinex (MSR) & MSR with Color Restoration (MSRCR)
Log
*
*
*
X
X
X
∑ Log ∑
),(F1 yx
),(F2 yx
),(F3 yx
),(I yx
MSR
-++
+
+
W1
W2
W3
*
Log
),(F yx
),(I yx
Log
∑ SSR-
+
CR
X
X
Gain/
OffsetMSRCR
Log
α
MSR
CRF
2019-04-10
8
02 Retinex Algorithm
∑ Multi-Scale Retinex (MSR) & MSR with Color Restoration (MSRCR)
N
n
n
N
n
nnMSR WRWRii
11
1,
2019-04-10
9
02 Retinex Algorithm
∑ MSRCR
linearity-nonofstrengththeControls:
ConstantGain:
ueOffset valGain :b
ConstantGain:G
]),(*),([),( byxRyxCGyxRii MSRiMSRCR
S
i
iii yxIyxIyxC1
),(log)],(log[),(
2019-04-10
03 Experiment Results
∑ Backlight Compensation
Origin
MSR
SSR Gamma Correction
sigma = 250 Gamma = 2.2
MSRCR
sigma =15, 80, 250
G = 130, b = 1
sigma =15, 80, 250
MMSRCR
sigma =80, 250
G = 130, b = 1
2019-04-10
03 Experiment Results
∑ Backlight Compensation
Origin
MSR
SSR Gamma Correctionsigma = 250 Gamma = 2.2
sigma =15, 80, 250
G = 130, b = 1
sigma =15, 80, 250 sigma =80, 250
G = 130, b = 1
MSRCR MMSRCR
2019-04-10
03 Experiment Results
∑ Backlight Compensation
Origin MSRSSR Gamma Correction
sigma = 250Gamma = 2.2
MSRCRsigma =15, 80, 250
G = 200, b = -20, α=100, β=40
sigma =15, 80, 250
MMSRCRsigma =80, 250
G = 200, b = 30
2019-04-10
04 Conclusion
13
SSR 보다는 MSR, MSRCR의성능이더좋음을확인.
처리속도가훨씬오래걸림.
실시간적용은힘들것으로예상.
http://dragon.larc.nasa.gov/retinex/pao/news/
Q & A