|
|
|
Image Enhancement
It increases distinction between features in a scene
1. Contrast manipulation
Gray-level threshold
- segmenting an image into
two classes - binary mask
- additional processing
can be applied to each segment
Level slicing
- dividing the histogram of DNs into several slices
- extensively used in thermal
images to show temp ranges
Contrast stretching
- expanding a narrow range
of DNs to a full range
DN - Min
Linear stretch: DN' = (--------------)
255
Max - Min
- advantage: simple computation
and efficient storage
- disadvantage: rare and frequent
values have the same amount of levels
Histogram-equalized stretch
- stretching based on frequency
of occurrence
- frequently occurred DNs
have more display levels
- special stretch
2. Spatial feature manipulation
Spatial filtering
- Low pass filters emphasize
low frequency features
- computing the average values of moving
windows
- High pass filter emphasizes
local details
- subtracting a low-pass
filter from the original image
Convolution
- a moving kernel with a
weighting factor for each pixel
Edge enhancement
- adding back the original
image to the high frequency image component
- preserving both the original
and the high frequency features
Directional first differencing
- displaying the differences
in gray levels of adjacent pixels
- the direction can be horizontal, vertical, or diagonal
- it is necessary to add a constant to the difference for display purposes
Contrast
stretching is needed for all feature manipulations
3. Multi-image manipulation
Spectral ratioing
- a ratio of two bands (with
great difference in reflectance)
- useful to eliminate effects of illumination
differences
- necessary to stretch the
resultant values to a full
range of DN after ratioing
Hybrid color ratio composite
- problem: different features of
similar ratio may appear identical
- solution: combining 2 ratio bands + 1 original
band to restore the absolute DN
Principle component transformation
- to reduce redundancy in multi-spectral
data
The transformation:
DNI = a11DNA + a12DNB
DNII = a21DNA + a22DNB
DNI, DNII = DNs in new component
images
DNA, DNB = DNs in the original
images
a11, a12, a21, a22 = coefficients
for the transformation
- after the axes rotation,
the original n bands images are
converted into n principle components images
- the first component (PC1)
image contains the largest
percentage of the total scene variance (90%±)
- the second component (PC2)
contains the largest % of the
remaining variance
- each succeeding component
contains decreasing amount of
variance
- successive components are
orthogonal; they are not
correlated to each other
- PCs can be used as new bands for image classification
- PCA is scene specific
Kauth-Thomas tasseled cap
- an orthogonal transformation
- the 4 MSS bands can be
converted into 4 new bands:
brightness,
greenness, yellow stuff, and non-such
- the first two indices contain the most info (90+%)
- brightness is related to bare soils
- greenness is related to the amount of green vegetation
- the 6 TM bands can be converted
into a 3-D space:
plane of soils,
plane of vegetation, and a transition zone
- a third feature, wetness
- the K-T transformation
is transferable between scenes
Intensity-Hue-Saturation transformation (IHS)
transform the RGB space into the IHS
space for
color enhancement
- intensity: brightness, hue: color, saturation: purity
- the hexcone model projects the RGB cube to a plane, resulting in a hexagon
- the plane is perpendicular
to the gray line and tangent to
the cube at the "white" corner
- intensity = distance along the gray line from the black point to any given hexagonal projection
- hue = angle around the hexagon
- saturation = distance from the gray point at the center of the hexagon
- I,H,S = f(R,G,B), I' = f(I+Ipan),
H' = f(H+Hpan), S' = f(S+Span)
R',G',B' = f(I',H',S')
- IHS is flexible in image
enhancement
4. Reading chpt 7