lecture #14 - case western reserve university
TRANSCRIPT
![Page 1: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/1.jpg)
EECS490: Digital Image Processing
Lecture #14
• Color Gradient & Color Edges
• Noise in Color Images
• Image Degradation & Restoration
• Noise
• Noise Filters - mean, -trimmed mean,
ordering, contraharmonic
![Page 2: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/2.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Color Edges
Color edges aligned
Color edges NOTaligned
RED GREEN BLUE COLOR
If we simplyadded themagnitudes ofthe gradientsthey would bethe same at thecenter.Intuitively theycannot be thesame.
![Page 3: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/3.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
(Vector) Color Gradient
The maximum rate of change of a color vector c(x,y) at
(x,y) is given by
in the direction
NOTE: This expression gives two directions. One is the direction
of the maximum of F; the other is the direction of the minimum.
F ( ) =1
2gxx + gyy( ) + gxx gyy( )cos2 + 2gxy sin2
x, y( ) =1
2Tan 1 2gxy
gxx gyy
S.D.Zenzo, “A Note on the Gradient of a Multi-Image,” Computer Vision, Graphicsand Image Processing, Vol. 33, pp.116-125, 1986.
See Section 6.6 and p. 563-564 of GWE. Implemented as colorgrad.
![Page 4: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/4.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
(Vector) Color Gradient
Let be the unit vectors along the RGB axes of a
RGB color space. Define
Further define
u =R
xr +
G
xg +
B
xb
v =R
yr +
G
yg +
B
yb
gxx = uiu = uTu =R
x
2
+G
x
2
+B
x
2
gyy = viv = vT v =
R
y
2
+G
y
2
+B
y
2
gxy = uiv = uTv =R
x
R
y+
G
x
G
y+
B
x
B
y
r, g, b
![Page 5: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/5.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Color Edges
Original image
Vector color gradient
Compute gradient ineach color image andadd together.
Difference betweengradient images.
![Page 6: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/6.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Color Edges
![Page 7: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/7.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Noise in RGB Color ImageAdditive Gaussian noise (mean=0, variance=800) in EACH color component
Noisy RED
Noisy BLUE
Noisy GREEN
Noisy RGB image
![Page 8: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/8.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Noise in HSI Color Image
I looks clean since1/3*(R+G+B) averagesnoise
Significantly degraded due to non-linearity ofcosine and min in HSI transformations
![Page 9: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/9.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Noise in Color Image
RGB image withsalt&pepper noise ingreen component
HUE
SATURATION INTENSITY
Noise spreads to allHSI components.
![Page 10: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/10.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods
Noise in Compression
Original image
Original image compressed anddecompressed using JPEG 2000.Slight blurring due to lossinherent in compressiontechnique.
JPEG 2000 uses a colortransformation similar to thatused by the Y’CbCr (luminance-blue chroma-red chroma) colorTV standard and waveletcompression.
![Page 11: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/11.jpg)
EECS490: Digital Image Processing
© 2002 R. C. Gonzalez & R. E. Woods 1999-2007 by Richard Alan Peters II
![Page 12: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/12.jpg)
EECS490: Digital Image Processing
Image Restoration
© 2002 R. C. Gonzalez & R. E. Woods
Limits us tocertain types ofdegradation
Additive noise
g x, y( ) = h x, y( ) f x, y( ) + x, y( )
G u,v( ) = H u,v( )F u,v( ) + N u,v( )
![Page 13: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/13.jpg)
EECS490: Digital Image Processing
Noise PDFs
© 2002 R. C. Gonzalez & R. E. Woods
Used forskewedhistograms
![Page 14: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/14.jpg)
EECS490: Digital Image Processing
Noise PDFs
© 2002 R. C. Gonzalez & R. E. Woods
Gaussian
p z( ) =1
2e
z μ( )2
2 2
p z( ) =2
bz a( )e
z a( )2
b z a
0 z < a
p z( ) =
abzb+1
b 1( )!z a( )e az z 0
0 z < 0
Rayleigh Gamma
Exponential
p z( ) =1
b aa z b
0 otherwise
p z( ) =
pa z = a
pb z = b
0 otherwise
Uniform(white)
Impulse(salt & pepper)
p z( ) =ae az z 0
0 z < 0
![Page 15: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/15.jpg)
EECS490: Digital Image Processing
Test Image
© 2002 R. C. Gonzalez & R. E. Woods
![Page 16: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/16.jpg)
EECS490: Digital Image Processing
Noisy Images
© 2002 R. C. Gonzalez & R. E. Woods
Usually hard to identify type of noise in spatial domain
Test Image + Noise
Noise spectra adjusted so they just overlap
![Page 17: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/17.jpg)
EECS490: Digital Image Processing
Noisy Images
© 2002 R. C. Gonzalez & R. E. Woods
Test Image + Noise
![Page 18: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/18.jpg)
EECS490: Digital Image Processing
Noisy Images
© 2002 R. C. Gonzalez & R. E. Woods
Note regular pattern ofnoise in image.
![Page 19: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/19.jpg)
EECS490: Digital Image Processing
Noise Characterization
© 2002 R. C. Gonzalez & R. E. Woods
Compute mean and variance for histograms and relate them to distribution parameters.
μz = zi p zi( )zi
z = zi μz( )2p zi( )
zi
zi is the gray level and p(zi) is the histogram
Test strips
![Page 20: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/20.jpg)
EECS490: Digital Image Processing
Mean Filters
© 2002 R. C. Gonzalez & R. E. Woods
f x, y( ) =1
mng s,t( )
(s,t ) Sxy
f x, y( ) = g s,t( )1
mn
(s,t ) Sxy
Arithmetic mean filter Geometric mean filter
Geometric mean filtergives smoothingcomparable toarithmetic mean filterwithout losing as muchdetail.
Original image Image + Gaussian noise
![Page 21: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/21.jpg)
EECS490: Digital Image Processing
Contraharmonic Filters
© 2002 R. C. Gonzalez & R. E. Woods
f x, y( ) =
g s,t( )Q+1
(s,t ) Sxy
g s,t( )Q
(s,t ) Sxy
Contraharmonic(Q=1.5)eliminates pepper noise
Contraharmonic(Q=-1.5)eliminates saltnoise
Contraharmonic filterreduces to arithmeticmean filter if Q=1
Pepper noise(random 0’s)
Salt noise(random 1’s)
Contraharmonic filter can reduce salt or pepper noise but not both at the same time.
![Page 22: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/22.jpg)
EECS490: Digital Image Processing
Contraharmonic Filters
© 2002 R. C. Gonzalez & R. E. Woods
Using the wrong sign in a contraharmonic filter will significantly degrade the image.
![Page 23: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/23.jpg)
EECS490: Digital Image Processing
Order Statistics Filter
© 2002 R. C. Gonzalez & R. E. Woods
Order statistics filters are
based upon ordering
(ranking) pixels in a
neighborhood.
The median filter selects the
middle element in the list.
f x, y( ) = median(s,t ) Sxy
g s,t( ){ }
![Page 24: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/24.jpg)
EECS490: Digital Image Processing
Order Statistics Filters
© 2002 R. C. Gonzalez & R. E. Woods
Other order statistics filters
Median works best for impulse
noise.
Max works best for finding
bright points.
Min works best for finding dark
points.
Midpoint works best for
Gaussian or uniform noise
f x, y( ) = median(s,t ) Sxy
g s,t( ){ }
f x, y( ) = max(s,t ) Sxy
g s,t( ){ }
f x, y( ) = min(s,t ) Sxy
g s,t( ){ }
f x, y( ) =1
2min(s,t ) Sxy
g s,t( ){ } + max(s,t ) Sxy
g s,t( ){ }
![Page 25: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/25.jpg)
EECS490: Digital Image Processing
Image Restoration
© 2002 R. C. Gonzalez & R. E. Woods
Max filter Min filter
Max and min filters do a reasonable job of removing impulsive noise but:• Max filter removed dark pixels from the borders of dark objects• Min filter removed light pixels from the borders of light objects
Notedifferencesbetween images
![Page 26: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/26.jpg)
EECS490: Digital Image Processing
Yet Another Order Statistics Filter
© 2002 R. C. Gonzalez & R. E. Woods
Alpha-trimmed mean filter
removes the d/2 highest and
d/2 lowest intensity values.
The average of these
remaining mn-d values is
called an alpha-trimmed
mean filter.
f x, y( ) =1
mn dgr s,t( )
(s,t ) Sxy
![Page 27: Lecture #14 - Case Western Reserve University](https://reader031.vdocuments.mx/reader031/viewer/2022011723/61d2f148c25e1c52b230a4f9/html5/thumbnails/27.jpg)
EECS490: Digital Image Processing
Image Restoration
© 2002 R. C. Gonzalez & R. E. Woods
Image+white noise
5x5arithmeticmeanfilter
Image+white noise+salt&peppernoise
5x5geometricmeanfilter
5x5medianfilter
5x5 alpha-trimmed(d=5) mean filter
Approaches performanceof median filter as dincreases but alsosmooths