BCS of
Images and
Video
J. E. Fowler
CS Overview
Images
Video
Multiview
Perspectives
Outline
1
Compressed Sensing (CS)
Overview of CS
Projected Landweber (PL) Recovery
2
CS for Images
Acquisition
Block-Based CS (BCS)
BCS-SPL
Results
3
CS for Video
CS for Video
Motion-Compensated BCS-SPL (MC-BCS-SPL)
Results
4
CS for Multiview Image Sets
Disparity-Compensated BCS-SPL (DC-BCS-SPL)
Results
5
Perspectives
BCS of
Images and
Video
J. E. Fowler
CS Overview
Overview of CS
PL Recovery
Images
Video
Multiview
Perspectives
Compressed Sensing (CS)
What is CS?
Emerging mathematical paradigm permitting:
Sampling at sub-Nyquist rates via linear projection
onto a measurement basis of lower dimension
Exact reconstruction when signal is sparse in some
transform domain
Approximate reconstruction when signal is
compressible in some transform domain
Random m easurem ent matrix works universally for
all signals with high probability
Also know as: compressive sensing, compressive
sampling
BCS of
Images and
Video
J. E. Fowler
CS Overview
Overview of CS
PL Recovery
Images
Video
Multiview
Perspectives
CS Overview
Goal
Recover vector x ∈ ℜ
N
from
y = Φx ∈ ℜ
M
Φ: M × N measurement mat rix, M ≪ N
Usually, Φ is a random matrix
Subsampling rat e, or subrate, is M/N
The m easurem ent process Φx is accomplished within
sensing device:
x is acquired and simultaneously reduced in
dimension
BCS of
Images and
Video
J. E. Fowler
CS Overview
Overview of CS
PL Recovery
Images
Video
Multiview
Perspectives
CS Overview
Fundamental Tenet of CS
Recovery is exact i f x is sufficiently sparse:
L-sparsity: only L coefficients of
ˇ
x = Ψx
are nonzero for some transform Ψ
Approximate Recovery
Real-world signals—often not sparse but compressible:
|
ˇ
x
n
| < Cn
−1/p
where p ≤ 1, C < ∞, and
ˇ
x
n
are sor ted coefficients of
ˇ
x
Recovery is close to L-sparse approximation to
ˇ
x
- 1
- 2
- 3
- 4
前往页