sparse image representation via combined transforms - Convex ...
sparse image representation via combined transforms - Convex ... sparse image representation via combined transforms - Convex ...
68 CHAPTER 3. IMAGE TRANSFORMS AND IMAGE FEATURES In the remainder of this subsection, we introduce four transforms: Gabor, chirplet, cosine packets and wavelet packets. They are selected because of their importance in the literature. (Of course the selection might be biased by personal preference.) Gabor Analysis The Gabor transform was first introduced in 1946 [69]. A contemporary description of the Gabor transform can be found in many books, for example, [68] and [100]. For a 1-D continuous function f(t) and a continuous time variable t, the Gabor transform is G(m, n) = ∫ +∞ −∞ f(t) · h(t − mT )e −iω 0nt dt, for integer m, n ∈ Z, (3.35) where h is a function with finite support, also called a window function. The constants T and ω 0 are sampling rates for time and frequency. A function b m,n (t) =h(t − mT )e −iω 0nt (3.36) is called a basis function of Gabor analysis. (Of course we assume that the set {b m,n : m, n ∈ Z} does form a basis.) Function b m,n (t) should be both time and frequency localized. A basis function of the Gabor transform is a shifted version of one initializing function. The shifting is operated in both time (by mT ) and frequency (by nω 0 ) domains, so the Gabor transform can be compared to partitioning the TFP into rectangles that have the same width and height. Figure 3.7 (c) depicts an idealized tiling corresponding to Gabor analysis. Choices of T and ω 0 determine the resolution of a partitioning. The choice of the window function h, which appears in both (3.35) and (3.36), determines how concentrated (in both time and frequency) a basis function b m,n is. Two theoretical results are noteworthy: • If a product T · ω 0 is greater than 2π, then a set {h(t − mT )e −iω0nt : m, n ∈ Z} can never be a frame for L 2 (R) (Daubechies [34], Section 4.1). • When T · ω 0 =2π, ifaset{h(t − mT )e −iω0nt : m, n ∈ Z} isaframeforL 2 (R), then either ∫ x 2 |g(x)| 2 dx = ∞ or ∫ ξ 2 |ĝ(ξ)| 2 dξ = ∞ (Balian-Low [34]). Gabor analysis is a powerful tool. It has become a foundation for the joint time frequency analysis.
3.4. OTHER TRANSFORMS 69 Chirplets A chirp is a signal whose instantaneous frequency is a linear function of the time. A detailed and concise definition of instantaneous frequency is in Flandrin [68]. A chirplet can be viewed as a piece of a chirp, or equivalently a windowized version of a chirp. The following is an example of a chirplet function: h(t − nT )e iq(t) , (3.37) where h is still a window function, T is the time sampling rate, and q(t) is a quadratic polynomial of t. An example of h is the Gaussian function h(t) =e −t2 . Note h is a real function. Suppose the general form for the quadratic polynomial is q(t) =a 2 t 2 + a 1 t + a 0 . The instantaneous frequency is (in this case) q ′ (t) =2a 2 t + a 1 , which is linear in t. On the TFP, a support of a chirplet is a needle-like atom. Figure 3.7(e) shows a chirplet. An early reference on this subject is [109]. A more general but complicated way to formulate chirplets is described in [103]. An idealistic description of the chirplet formulation in [103] is that we consider the chirplet as a quadrilateral on the time frequency plane. Since each corner of the quadrilateral has two degrees of freedom, the degrees of freedom under this formulation can go up to 8. A signal associated with this quadrilateral is actually a transform of a signal associated with a rectangle in TFP; for example, the original signal could be a Gabor basis function. There are eight allowed transforms: (1) translation in time, (2) dilation in time, (3) translation in frequency, also called modulation, (4) dilation in frequency, (5) shear in time (Fourier transformation, followed by a multiplication by a chirp, then an inverse Fourier transformation), (6) shear in frequency (multiplication by a chirp), (7) perspective projection in the time domain, and finally (8) perspective projection in the frequency domain. Generally, a chirplet system is overcomplete. The following is about how to build a chirplet-like orthonormal basis. It is instructive to see that a new orthonormal basis can be built out of an existing orthonormal basis. The following idea is described in [7]. They introduced an operator called axis warping: iff(x) is a function in L 2 (R), the axis warping of function f(x) is (W c f)(x) =|c| 1/2 |x| (c−1)/2 f(|x| c sign(x)),
- Page 45 and 46: 2.2. SPARSITY AND COMPRESSION 17 wi
- Page 47 and 48: 2.2. SPARSITY AND COMPRESSION 19 lo
- Page 49 and 50: 2.3. DISCUSSION 21 tail compact. Th
- Page 51 and 52: 2.4. PROOF 23 The index l does not
- Page 53 and 54: Chapter 3 Image Transforms and Imag
- Page 55 and 56: 27 Some of the figures show the bas
- Page 57 and 58: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 59 and 60: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 61 and 62: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 63 and 64: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 65 and 66: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 67 and 68: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 69 and 70: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 71 and 72: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 73 and 74: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 75 and 76: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 77 and 78: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 79 and 80: 3.1. DCT AND HOMOGENEOUS COMPONENTS
- Page 81 and 82: 3.2. WAVELETS AND POINT SINGULARITI
- Page 83 and 84: 3.2. WAVELETS AND POINT SINGULARITI
- Page 85 and 86: 3.2. WAVELETS AND POINT SINGULARITI
- Page 87 and 88: 3.2. WAVELETS AND POINT SINGULARITI
- Page 89 and 90: 3.2. WAVELETS AND POINT SINGULARITI
- Page 91 and 92: 3.2. WAVELETS AND POINT SINGULARITI
- Page 93 and 94: 3.3. EDGELETS AND LINEAR SINGULARIT
- Page 95: 3.4. OTHER TRANSFORMS 67 uncertaint
- Page 99 and 100: 3.4. OTHER TRANSFORMS 71 Folding. A
- Page 101 and 102: 3.4. OTHER TRANSFORMS 73 We can app
- Page 103 and 104: 3.5. DISCUSSION 75 give only a few
- Page 105 and 106: 3.7. PROOFS 77 the ijth component o
- Page 107 and 108: 3.7. PROOFS 79 Similarly, we have [
- Page 109 and 110: Chapter 4 Combined Image Representa
- Page 111 and 112: 4.2. SPARSE DECOMPOSITION 83 interi
- Page 113 and 114: 4.3. MINIMUM l 1 NORM SOLUTION 85 l
- Page 115 and 116: 4.4. LAGRANGE MULTIPLIERS 87 ρ( x
- Page 117 and 118: 4.5. HOW TO CHOOSE ρ AND λ 89 3 (
- Page 119 and 120: 4.6. HOMOTOPY 91 A way to interpret
- Page 121 and 122: 4.7. NEWTON DIRECTION 93 4.7 Newton
- Page 123 and 124: 4.9. ITERATIVE METHODS 95 1. Avoidi
- Page 125 and 126: 4.11. DISCUSSION 97 ρ(β) =‖β
- Page 127 and 128: 4.12. PROOFS 99 4.12.2 Proof of The
- Page 129 and 130: 4.12. PROOFS 101 case of (4.16). Co
- Page 131 and 132: Chapter 5 Iterative Methods This ch
- Page 133 and 134: 5.1. OVERVIEW 105 the k-th iteratio
- Page 135 and 136: 5.1. OVERVIEW 107 5.1.4 Preconditio
- Page 137 and 138: 5.2. LSQR 109 among all the block d
- Page 139 and 140: 5.2. LSQR 111 5.2.3 Algorithm LSQR
- Page 141 and 142: 5.3. MINRES 113 2. For k =1, 2,...,
- Page 143 and 144: 5.3. MINRES 115 using the precondit
- Page 145 and 146: 5.4. DISCUSSION 117 From (I + S 1 )
3.4. OTHER TRANSFORMS 69<br />
Chirplets<br />
A chirp is a signal whose instantaneous frequency is a linear function of the time. A<br />
detailed and concise definition of instantaneous frequency is in Flandrin [68]. A chirplet<br />
can be viewed as a piece of a chirp, or equivalently a windowized version of a chirp. The<br />
following is an example of a chirplet function:<br />
h(t − nT )e iq(t) , (3.37)<br />
where h is still a window function, T is the time sampling rate, and q(t) is a quadratic<br />
polynomial of t. An example of h is the Gaussian function h(t) =e −t2 . Note h is a real<br />
function. Suppose the general form for the quadratic polynomial is q(t) =a 2 t 2 + a 1 t + a 0 .<br />
The instantaneous frequency is (in this case) q ′ (t) =2a 2 t + a 1 , which is linear in t.<br />
On the TFP, a support of a chirplet is a needle-like atom. Figure 3.7(e) shows a chirplet.<br />
An early reference on this subject is [109]. A more general but complicated way to formulate<br />
chirplets is described in [103]. An idealistic description of the chirplet formulation in [103]<br />
is that we consider the chirplet as a quadrilateral on the time frequency plane. Since<br />
each corner of the quadrilateral has two degrees of freedom, the degrees of freedom under<br />
this formulation can go up to 8. A signal associated with this quadrilateral is actually a<br />
transform of a signal associated with a rectangle in TFP; for example, the original signal<br />
could be a Gabor basis function. There are eight allowed <strong>transforms</strong>: (1) translation in<br />
time, (2) dilation in time, (3) translation in frequency, also called modulation, (4) dilation<br />
in frequency, (5) shear in time (Fourier transformation, followed by a multiplication by a<br />
chirp, then an inverse Fourier transformation), (6) shear in frequency (multiplication by a<br />
chirp), (7) perspective projection in the time domain, and finally (8) perspective projection<br />
in the frequency domain.<br />
Generally, a chirplet system is overcomplete. The following is about how to build a<br />
chirplet-like orthonormal basis. It is instructive to see that a new orthonormal basis can<br />
be built out of an existing orthonormal basis. The following idea is described in [7]. They<br />
introduced an operator called axis warping: iff(x) is a function in L 2 (R), the axis warping<br />
of function f(x) is<br />
(W c f)(x) =|c| 1/2 |x| (c−1)/2 f(|x| c sign(x)),