Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
Experimenting with Multidimensional Wavelet
Transformations
Tarık Arıcı and Buğra Gedik
Outline of Project Goals
Writing discrete wavelet transformation and inverse
transformation wrappers (in Matlab) to handle multidimensional data; possible uses include:
2D Images, 3D turbulence data or multi-attribute
sensor readings
Using wavelets in some example applications
Lossy compression, De-noising for images, Selfsimilarity analysis
Studying the phases of the wavelet filters (that delays
the wavelet smoothes) and approximately computing
the delay amount using DSP methods
Using this on Mammogram reconstruction
Possible uses of Bayesian? (not done)
DWTR / IDWTR wrappers
Assume D dimensions
Perform D sweeps, one across
each dimension, making recursive smoothes
calls for each D-1 dimensional slice
Top level recursive calls go D-1
levels deep before calling the 1
dimensional wavelet transformation
functions
7 detail groups
As a result 2^D-1 detail groups and
a single smooth group is
constructed for each level of
transformation
smoothes
3 detail groups
Example Applications: Lossy
Compression
Example Applications: De-noising
Example Applications: Self-similarity
Analysis
Calculate the means of the detail squares for each
level and plot their log as a function of level
If the line is linear, then there is self-similarity
Brownian motion is self-similar, Random data (of
course) is not
Mammogram Reconstruction
Original Image
after wavelet
interpolation
after fixing
delay problem
Assume all details are zero
Perform inverse wavelet transformation
Possible use of Bayesian Methods:
Model missing details using a Bayesian approach
DSP Perspective: Problems Related
with Non-zero Phase Filtering
Filtering in time domain is multiplication in frequency
domain
X[n]
h[n]
y[n]
Phase(Y(f)) = Phase(H(f))+Phase(X(f))
Non-zero Phase Filtering
cos(2pf0t+f) = cos(2pf0(t+f/(2pf0))
= f/(2pf0)
td is constant if f is a linear function of
frequency
td
Therefore, wavelet filters should be
(approximately) linear phase filters
Symmetric filters have linear phase
Ex: {1, 1} (Haar), {1, 2, 1}
Least Asymmetric (LA) Wavelet Filters
Choose filter coefficients:
s.t. min |f(f) – 2pfv|
v=
-L/2+1, if L =8,12,16,20
-L/2, if L =10, 18
-L/2+2, if L =14
LA(8) and LA(12) works best.
The End!
Thanks!!!