Fractal Dimension for Multiband Images

advertisement
Segmenting multi bands
images by color and
texture
Eldman O. Nunes
-
IC - UFF
Aura Conci
Introduction
•Use of fractals and image multiespectral
bands to characterize texture.
•Considering inter-relation among bands the
image FD є [ 0 , number of bands + 2] .
•Improve the possibilies of usual false color
segmentations (assigning satellite bands to RGB
color). It is not now limited to 3 band.
• The color sensations noticed by humans are
combination of the intensities received by 3
types of cells cones.
• Combination of the 3 primary colors produces
the others
• In the video: R=700 nm, G = 546,1 nm, B=435,1
nm.
Digital images
Monocromatic : one color channel or one band.
• binary image:
each pixel only
0 or 1 values.
• intensity level (grey level):
each pixel one value
from 0 to 255.
• Multiband images: n band value for each pixel.
• examples:
»color images
»sattelite images
»medical images
color images
each pixel 3 values ( from 0 to 255 )
3 bands: Red - Green -Blue.
Band 1
Band 2
Band 4
Band 5
Band 3
Band 6
Band 7
example : a LandSat-7 image is a collection
of 7 images of same scene
sensor
TM
characteristics
HRV
AVHRR
spacial
resolution
30 m
120 m (Band 6)
20 m (Band 1 a 3)
10 m (Pan)
1.1 Km (nominal)
spectral
Bands
(micro
meters)
Band 1 - 0.45-0.52
Band 2 - 0.52-0.60
Band 3 - 0.63-0.69
Band 4 - 0.76-0.90
Band 5 - 1.55-1.75
Band 6 - 10.74-12.5
Band 7 - 2.08-2.35
Band 1 - 0.50-0.59
Band 2 - 0.61-0.68
Band 3 - 0.79-0.89
Pan - 0.51-0.73
Band
Band
Band
Band
Band
Radiometric
resolution
8 bits
8 bits (1-3)
6 bits (Pan)
10 bits
16 days
26 days
2 times a days
Temporal
resolution
1 - 0.58-0.68
2 - 0.725-1.1
3 - 3.55-3.93
4 - 10.30-11.30
5 - 11.50-12.50
Landsat 7 - Sensor TM
Channel
spectral band (um)
main applications
1
0.45 - 0.52
Differentiation between soil and
vegetation, conifers and deciduous trees
2
0.52 - 0.60
healthy vegetation
3
0.63 - 0.69
chlorophyll absortion, vegetation types
4
0.76 - 0.90
biomass , water bodies
5
1.55 - 1.75
penetrate smokes, snow
6
10.4 - 12.5
surface temperature from -100 to 150 C
7
2.08 - 2.35
hidrotermal map, buildings, soil trafficability
Multiespectral false color :
l , m, n Bands to Red, Green and Blue.
Band 4 (R), 5 (G), 3 (B)
Band 4 (R), 3 (G), 2 (B)
Textures
Texture is characterized by the repetition of a model on an
area.
Textons : size, format, color and orientation of the elements.
Textons can be repeated in an exact way or with small
variations on a same theme.
Texture 1
Texture 2
Fractal Geometry
• self similar sets
• fractal dimensions and measures used to classify textures
FD for binary image
• Box Counting Theorem - 2D images.
• For a set A, Nn(A) = number of boxes of side 1/2n
which interser the set A:
DF = lim n log Nn (A) / log 2n
n
Nn (A)
2n
log Nn (A)
log 2n
1
4
2
1,386
0,693
2
12
4
2,484
1,386
3
36
8
3,583
2,079
4
108
16
4,682
2,772
5
324
32
5,780
3,465
6
972
64
6,879
4,158
log Nn (A)
8
6
4
2
0
0
1
2
3
log (2n )
4
5
gray level images
• Box Counting Theorem extension for 3-dimensional object:
third coordinate represents the intensity of the pixel.
• DF between 2 e 3.
Blanket Dimension - Blanket Covering Method
The space is subdivided in cubes of sides SxSxS ’.
Nn(A) denotes the number of cubes intercept a blanket
covering the image: Nn =  nn (i,j)
On each grid
(i,j),
nn (i,j) = int ( ( max – min ) / s’ ) + 1
for multi-bands image
•a color R G B image is a subset of the 5-dimensional
space : N5). Each pixel is defined by: (x, y, r, g, b)
•FD of this images: values from 2 to 5.
Generalizing: d-cube
• points (0D), segments (1D), squares (2D), cubes (3D) and
• for a n-dimensional : n-cube (nD)
• But what is d-cubos , and how many d-cubes appear
in a divison of Nd space?
r
r
r
r
r
r
SEGMENTO
QUADRADO
CUBO
Sweep representation :
• n-cube as translational swepps of (n-1) cube
Generalizing: d-Cube Counting - DCC:
• the experimental determination of the fractal
dimension of images with multiple
channels;
• will imply in the recursive division of the N
space in d-cubes of size r;
• followed by the contagem of the numbers of
d-cubes that intercept the image.
the space N3 is divided
by 3-cubos of size 1/2n, and the number of 3cubos that intercept the image it is counted.
• monochrome images:
• color images: the space N5 is divided by 5cubos of the same size 1/2n, and the number of
5-cubos that intercept the image is counted.
• satellite images: the space Nd is divided by dcubes of size 1/2n and the number of d-cubes
that intercept the image is counted.
•
number of 1-cubes:
Nn
1-cubos = 2 1x n, where n is the number of divisions.
•
number of 2-cubes:
Nn
2-cubos = 2 2x n, where n is the number of divisions.
•
number of 3-cubos: Nn
3-cubos = 2 3x n, where n is the number of divisions.
•
Generalizing, the number of identical d-cube:
Nn
d-cubes = 2 d x n, where d is the space dimension and n it is the
number of divisions.
Then FD of d-dimensional images can be obtained by:
DFn = log (Nn,d-cubo) /log (2n )
Results

binary images

gray scale

colored images

satellite images
CDC invariance to resolution (FD  3,465)
CDC invariance on colors reflection (second
image) and affine transformations (FD  3,465)
CDC invariance to band combinations(FD  3,465)
: RGB (4-5-6, 4-6-5, 5-4-6, 5-6-4, 6-4-5, 6-5-4)
Mosaic of textures: original x CDC
Segmentation result: same color means same texture.
comparison:
original
SPRING
-
SEGWIN
CDC
Region on the city of Patriocínio - MG
(from Landsat 5-TM, 5-4-3 spectral band to RGB)
Segmentation results by CDC
Download