Академический Документы
Профессиональный Документы
Культура Документы
Application to images
Vclav Hlav
Eigen-analysis
2/26
This linear space has some natural orthogonal basis vectors which allow
Ax = x ,
standpoint.
Later, we will develop a statistical view based on covariance matrices and
This system has a unique solution if and only if the rank of the matrix A is
Matrices A and B with real or complex entries are called similar if there
Let I be the unitary matrix having values 1 only on the main diagonal and
zeros elsewhere.
The polynomial of degree n given as det(A I) is called the
characteristic polynomial.
Then the eigen-equation A x = x holds if det(A I) = 0.
form
i 1 0
J1 0 . .
.
0 i . 0
. .
, where J i are Jordan blocks
,
0 0
. . . 1
0 Jp
0 0 i
If the eigen-value is not multiple then the Jordan block degenerates to the
eigen-value itself.
Least-square approximation
10/26
error.
Principal component analysis, introduction
11/26
(dimensions). Usually N M .
The aim: to reduce the dimensionality of the data so that each observation
because N may be very large: this is in fact good, since many observations
imply better statistics.
Data normalization is needed first
13/26
This procedure is not applied to the raw data R but to normalized data X
as follows.
The raw observed data is arranged in a matrix R and the empirical mean is
X = R ue .
Derivation, M -dimensional case (2)
14/26
N L N
!
1 X X 1 X
2
= 2
|xn| b>
i xn x >
n bi ,
N n=1 i=1
N n=1
L
X N
X
b>
i cov(x) bi , where cov(x) = xn x >
n ,
i=1 n=1
smallest) and the associated eigen-vectors taken as the basis vectors that
provide the maximum we seek.
In the data approximation, dimensions corresponding to the smallest
L
X M
X
2 = trace cov(x)
i = i ,
i=1 i=L+1
under-constrained.
PCA is still applicable.
}
}
... ~
~ ...
...
one PCA
represented image
= q1 + q2 + q3 + q4
Reconstruction fidelity, 4 components
21/26
Reconstruction fidelity, original
22/26
PCA drawbacks, the images case
23/26
Discriminative representation
Does not allow partial reconstruction.
the manifold.
Example (courtesy Thomas Brox):
The 100 100 image of the number 3 shifted and rotated, i.e. there are
only 3 degrees of variations.
Subspace methods explore the fact that data (images) can be represented in a
subspace of the original vector space in which data live.