Date post: | 01-Apr-2015 |
Category: |
Documents |
Upload: | emilie-justin |
View: | 221 times |
Download: | 6 times |
ter Haar Romeny, FEV 2011
Regularization
and scale-space
ter Haar Romeny, FEV 2011
Regularization is the technique to make data behave well when an operator is applied to them. A small variation of the input data should lead to small change in the output data.
Differentiation is a notorious function with 'bad behaviour'.
2 4 6 8 10 12
4
3
2
1
1
2
3
10 20 30
1
0.5
0.5
2 1 1 2
0.2
0.4
0.6
0.8
2 1 1 2 3 4 5
0.2
0.4
0.6
0.8
1
Some functions that can not be differentiated.
ter Haar Romeny, FEV 2011
In mathematical terms it is said that the operation of differentiation is ill-posed, the opposite of well-posed. Jacques Hadamard (1865–1963) stated the conditions for well-posedness:
• The solution must exist;• The solution must be uniquely determined;• The solution must depend continuously on the initial or boundary data.
In other words, the solution should be stable.
ter Haar Romeny, FEV 2011
• smoothing the data, convolution with some extended kernel, like a 'running average filter' or the Gaussian;• interpolation, by a polynomial (multidimensional) function;• energy minimization, of a cost function under constraints• fitting a function to the data (e.g. splines). The cubic splines are named so because they fit to third order;• graduated convexity [Blake1987];• deformable templates ('snakes') [McInerney1996];• thin plates splines [Bookstein1989];• Tikhonov regularization.
ter Haar Romeny, FEV 2011
The formal mathematical method to solve the problems of ill-posed differentiation was given by Laurent Schwartz:
A regular tempered distribution associated with an image is defined by the action of a smooth test function on the image. TL
Lxx x
i1...inTL 1n
Lxi1...inx xThe derivative is:
ter Haar Romeny, FEV 2011
Fields Medal 1950 for his work on the theory of distributions. Schwartz has received a long list of prizes, medals and honours in addition to the Fields Medal. He received prizes from the Paris Academy of Sciences in 1955, 1964 and 1972. In 1972 he was elected a member of the Academy. He has been awarded honorary doctorates from many universities including Humboldt (1960), Brussels (1962), Lund (1981), Tel-Aviv (1981), Montreal (1985) and Athens (1993).
Laurent Schwartz (1915 - 2002)
ter Haar Romeny, FEV 2011
Mathematics Smooth test function
Computer vision Kernel, filter
Biological vision Receptive field
ter Haar Romeny, FEV 2011
Relation regularization - Gaussian scale-space
An essential result in scale-space theory was shown by Mads Nielsen (Copenhagen University). He proved that Tikhonov regularization is essentially equivalent to convolution with a Gaussian kernel.
Eg
f g2 x minimize this function for g, given the constraint that the derivative behaves well.
Eg
f g2 1gx2 xEuler-Lagrange:
ter Haar Romeny, FEV 2011
In the Fourier domain the expressions are easier:
d Egd g
2f g2 212g 0
f g 12g 0 g 1112 f
E g F
f g2 1 g x2
f g2 1 gx
2
f g2 1 2 gx2
ter Haar Romeny, FEV 2011
3 2 1 1 2 3
0.2
0.4
0.6
0.8
1.2
1
1 x21
1
In the spatial domain:
Filter proposed by Castan, 1990
ter Haar Romeny, FEV 2011
f g2 12g2 24g2
Including the second order derivative:
d Egd g
2f g2 212g 24g 0
g1
1 12 24f
4 2 2 4
0.2
0.3
0.4
0.5
0.6
Function proposed by Deriche (1987)
Etc.
E g
f g2 1 gx
2 2 gx x
2
ter Haar Romeny, FEV 2011
12 2 2
2 1
2 2
2
4 4
8
6 6
48
8 8
384
10 10
3840 O11
Taylor expansion of the Gaussian in the Fourier domain:
By recursion:
Tikhonov regularization is equivalent to Gaussian blurring
4 2 2 4
0.2
0.4
0.6
0.8
1
ter Haar Romeny, FEV 2011
121
000
121
horSobel
101
202
101
vertSobel
1 0 1
1
0
1
vert
hor
Neighbor
Neighbor
Older implementations are approximations of the Gaussian kernel:
1 1
1
1
vert
hor
Neighbor
Neighbor