-4-
why spatial processing
http://flickr.com/photos/alliwalk/3284897415/
-5-
roadmap for today
Application
Method
f ⎯T⎯
⎯→ g = TΝ ( f )
N (.)
f ( x, y ) , 1 ≤ x ≤ M ,1 ≤ y ≤ N g ( x, y ) , 1 ≤ x ≤ M ,1 ≤ y ≤ N
TN (.) : Spatial operator defined on a neighborhood N of a given pixel
N 0 ( x, y ) N 4 ( x, y ) N 8 ( x, y )
point processing mask/kernel
processing
-6-
outline
What and why
Spatial domain processing for image
enhancement
Intensity Transformation
Spatial Filtering
-7-
intensity transformation / point operation
Map a given gray or color level u to a
new level v
v
Memory-less, direction-less operation
output gray level
output at (x, y) only depend on the input
intensity at the same point
Pixels of the same intensity gets the same
transformation
Does not bring in new information, may
cause loss of information input gray level u
But can improve visual appearance or
make features easier to detect
-8-
intensity transformation / point operation
Two examples we already saw
Color space transformation
Scalar quantization
-9-
image negatives
the appearance of
photographic
negatives
Enhance white or
gray detail on dark
regions, esp. when
black areas are
dominant in size
-10-
basic intensity transform functions
monotonic, reversible
compress or stretch certain range of gray-levels
-11-
log transform
lena
stretch:
FFT(lena) u ∈ [0, .5] Æ
v ∈ [0, .59]
compress:
im = imread(‘lena.png’) u ∈ [.5, 1] Æ
a = abs(fftshift(fft2(double(im)))); v ∈ [.59, 1]
c = log(1+double(im)); c = range_normalize(c);
b = log(1+a); b=b/max(b(:));
-12-
-13-
power-law transformation
power-law response
functions in practice
CRT Intensity-to-voltage
function has
γ ≈ 1.8~2.5
Camera capturing distortion
with γc = 1.0-1.7
Similar device curves in
scanners, printers, …
power-law transformations are also useful for
general purpose contrast manipulation
-14-
gamma correction
make linear input appear linear on displays
method: calibration pattern + interactive adjustment
example calibration chart
-15-
effect of gamma on consumer photos
2.2 1/2.2
L0 L0 L0
-16-
what gamma to use?
γ >1
γ <1
?
-17-
more intensity transform
log, gamma … closed-form functions on [0,1]
can be more flexible
contrast stretching
-18-
intensity slicing
-19-
image bit-planes
-20-
slicing bitplanes
Depend on relative importance of bits
How much to slice depend on image content
Useful in image compression, e.g. JPEG2000
-21-
outline
What and why
Image enhancement
Spatial domain processing
Intensity Transformation
Intensity transformation functions (negative, log,
gamma), intensity and bit-place slicing, contrast
stretching
Histograms: equalization, matching, local
processing
Spatial Filtering
Filtering basics, smoothing filters, sharpening
filters, unsharp masking, laplacian
Combining spatial operations
-22-
gray-level image histogram
Represents the relative frequency of occurrence of the
various gray levels in the image
For each gray level, count the number of pixels having that level
Can group nearby levels to form a big bin & count #pixels in it
-23-
interpretations of histogram
if pixel values are i.i.d random
variables Æ histogram is an
estimate of the probability
distribution of the r.v.
“unbalanced” histograms do
not fully utilize the dynamic
range
Low contrast image: narrow
luminance range
Under-exposed image:
concentrating on the dark side
Over-exposed image:
concentrating on the bright
side
“balanced” histogram gives
more pleasant look and reveals
rich details
-24-
contrast stretching
Stretch the over-concentrated gray-levels
Piece-wise linear function, where the slope in the
stretching region is greater than 1.
β = T (α )
L-1
β2 s3
s2
β1 s1
0 α1 α 2 L-1 α
-25-
… in practice
intuition about a “good” image:
a “uniform” histogram spanning a large variety of gray tones
can we figure out a stretching function automatically?
-26-
histogram equalization
goal: map the each luminance level to a new value such that the
output image has approximately uniform distribution of gray levels
two desired properties
monotonic (non-decreasing) function: no value reversals
[0,1]Æ[0,1] : the output range being the same as the input range
pdf
1 1
cdf
o 1 o 1
-27-
histogram equalization
make
1
show that v follows uniform distribution, i.e.
o 1
-28-
implementing histogram equalization
u v v’
v = ∑ pu ( xi ) Rounding or
xi ≤u Uniform
quantization
pu(xi)
compute n( xi )
pu ( xi ) = L −1
for i = 0, ..., L - 1
histogram
∑ n( x )
i =0
i
u Only depend on the input
equalize v = ( L − 1) ∑ pu ( xi ) image histogram
xi = 0
L −1 u Fast to implement
or v= ∑ n( xi ) For u in discrete prob.
MN xi =0 distribution, the output v
round the will be approximately
output v' = round (v) uniform
-29-
a toy example
u
v = ( L − 1) ∑ pu ( xi )
xi = 0
-30-
a toy example
1.33 1.33 1
1.75 3.08 3
1.47 4.55 5
1.02 5.67 6
0.56 6.23 6
0.42 6.65 7
0.21 6.86 7
0.14 7.00 7
-31-
histogram equalization example
-32-
contrast-stretching vs. histogram equalization
v
γ
output gray level
α
a b
o input gray level u
function form
reversible? loss of information?
input/output?
automatic/interactive?
-33-
histogram matching
Histogram matching/specification
Want output v with specified p.d.f. pV(v)
Use a uniformly distributed random vairable
W as an intermediate step
W = FU(u) = FV(v) Æ V = F-1V (FU(u) )
Approximation in the intermediate step
needed for discrete r.v.
W1 = FU(u) , W2 = FV(v) Æ
take v s.t. its w2 is equal to or just above w1
-34-
histogram matching example
Histogram
equalized
-35-
local histogram processing
problem: global spatial processing not always desirable
solution: apply point-operations to a pixel
neighborhood with a sliding window
-36-
histogram equalization not always desirable
-37-
outline
What and why
Image enhancement
Spatial domain processing
Intensity Transformation
Intensity transformation functions (negative, log,
gamma), intensity and bit-place slicing, contrast
stretching
Histograms: equalization, matching, local
processing
Spatial Filtering
Filtering basics, smoothing filters, sharpening
filters, unsharp masking, laplacian
Combining spatial operations (sec. 3.7)
-38-
spatial filtering in image neighborhoods
-39-
kernel operator / filter masks
TN (.) = w(.)
f Spatial g
Filtering
kernel
a b
g (m, n) = ∑ ∑ w(i, j ) f (m + i, n + j )
i =− a j =−b
1≤ m ≤ M
1≤ n ≤ N
-40-
smoothing: image averaging
smoothing
operator
Low-pass filter, leads to
softened edges
-41-
spatial averaging can suppress noise
image with iid noise y(m,n) = x(m,n) + N(m,n)
UMCP ENEE408G Slides (created by M.Wu & R.Liu © 2002)
averaging
v(m,n) = (1/Nw) Σ x(m-k, n-l) + (1/Nw) Σ N(m-k, n-l)
Nw: number of pixels in the averaging window
Noise variance reduced by a factor of Nw
SNR improved by a factor of Nw
Window size is limited to avoid excessive blurring
-42-
smoothing operator of different sizes
original 3x3
5x5 9x9
15x15 35x35
-43-
directional smoothing
Problems with simple spatial averaging mask
UMCP ENEE408G Slides (created by M.Wu & R.Liu © 2002)
Edges get blurred
Improvement θ
Restrict smoothing to along edge direction
Avoid filtering across edges Wθ
Directional smoothing
Compute spatial average along several directions
Take the result from the direction giving the smallest changes
before & after filtering
Other solutions
Use more explicit edge detection and adapt filtering accordingly
-44-
non-linear smoothing operator
Median filtering
median value ξ over a small window of size Nw
nonlinear
median{ x(m) + y(m) } ≠ median{x(m)} + median{y(m)}
odd window size is commonly used
3x3, 5x5, 7x7
5-pixel “+”-shaped window
for even-sized windows take the average of two
middle values as output
Other order statistics: min, max, x-percentile …
-45-
median filter example
Median filtering
resilient to statistical outliers
incurs less blurring
simple to implement
iid noise
more at lecture 7, “image restoration”
-46-
image derivative and sharpening
-47-
edge and the first derivative
Edge: pixel locations of abrupt luminance
change
Spatial luminance gradient vector
a vector consists of partial derivatives along two
orthogonal directions
gradient gives the direction with highest rate of
luminance changes
Representing edge: edge intensity + directions
Detection Methods
prepare edge examples (templates) of different
intensities and directions, then find the best match
measure transitions along 2 orthogonal directions
-48-
edge detection operators
Image gradient:
⎡ ∂f ⎤
⎡Gx ⎤ ⎢ ∂x ⎥
∇f = ⎢ ⎥ = ⎢ ∂f ⎥
⎣G y ⎦ ⎢ ⎥
⎢⎣ ∂y ⎥⎦
∇f ≈ G x + G y
Robert’s
operator
Sobel’s
operator
-49-
edge detection examples
Roberts
Sobel
http://flickr.com/photos/reneemarie11/97326485
http://flickr.com/photos/valkyrie2112/2829322895
-50-
second derivative in 2D
Image Laplacian:
-51-
laplacian of roman ruins
http://flickr.com/photos/starfish235/388557119/
-52-
unsharp masking
Unsharp masking is an image manipulation technique for
increasing the apparent sharpness of photographic images.
The "unsharp" of the name derives from the fact that the technique
uses a blurred, or "unsharp", positive to create a "mask" of the
original image. The unsharped mask is then combined with the
negative, creating a resulting image sharper than the original.
Steps
Blur the image
Subtract the blurred
version from the original
(this is called the mask)
Add the “mask” to the
original
-53-
high-boost filtering
Avg.
-
+
+ f hb ( x, y ) = Af ( x, y ) − f lp ( x, y )
Unsharp mask:
high-boost with A=1
-54-
unsharp mask example
-55-
unsharp mask example
http://flickr.com/photos/dpgnashua/2274968238/
-56-
summary
Spatial transformation and filtering are popular
methods for image enhancement
Intensity Transformation
Intensity transformation functions (negative, log,
gamma), intensity and bit-place slicing, contrast
stretching
Histograms: equalization, matching, local processing
Spatial Filtering
smoothing filters, sharpening filters, unsharp masking,
laplacian
Combining spatial operations (sec. 3.7)
-57-
sharpen !
http://flickr.com/photos/t_schnitzlein/87607390/
-58-
