Laplacian Colormaps: a framework for structure-preserving color transformations
-
Upload
davide-eynard -
Category
Science
-
view
824 -
download
1
description
Transcript of Laplacian Colormaps: a framework for structure-preserving color transformations
Laplacian colormaps: a frameworkfor structure-preserving color transformations
Davide Eynard, Artiom Kovnatsky, Michael Bronstein
Institute of Computational Science, Faculty of InformaticsUniversity of Lugano, Switzerland
Eurographics, 8 April 2014
This research was supported by the ERC Starting Grant No. 307047 (COMET).
1 / 40
2 / 40
3 / 40
4 / 40
Color transformations
RGB source Luma
Standard color transformations may break image structure!
5 / 40
Color transformations
RGB source Luma
Standard color transformations may break image structure!
6 / 40
Color transformations
RGB source Luma Desired outcome
Standard color transformations may break image structure!
7 / 40
Image Laplacian
Input N ×M image with d colorchannels, column-stacked into anNM × d matrix X
Represented as graph with K vertices(e.g. superpixels) and weighted edges
K ×K adjacency matrix WX
wij = exp
(−δ2ij2σ2s
+‖x′ki − x′kj‖
22
2σ2r
)
K ×K Laplacian
LX = DX−WX, DX = diag(∑j 6=i
wij)
8 / 40
Image Laplacian
Input N ×M image with d colorchannels, column-stacked into anNM × d matrix X
Represented as graph with K vertices(e.g. superpixels) and weighted edges
K ×K adjacency matrix WX
wij = exp
(−δ2ij2σ2s
+‖x′ki − x′kj‖
22
2σ2r
)
K ×K Laplacian
LX = DX−WX, DX = diag(∑j 6=i
wij)
x′ki
x′kj
wij
9 / 40
Laplacians = structure descriptors
UTLXU = ΛX, VTLYV = ΛY
X u4 u5 u6 u7
Y v4 v5 v6 v7
Similar structure ⇐⇒ similar Laplacian eigenvectors
10 / 40
Laplacians = structure descriptors
UTLXU = ΛX, VTLYV = ΛY
X u4 u5 u6 u7
Y v4 v5 v6 v7
Similar structure ⇐⇒ similar Laplacian eigenvectors
Ideally, two Laplacians are jointly diagonalizable (iff theycommute): there exists a joint eigenbasis U = U = V
11 / 40
Laplacians = structure descriptors
X u2 u3 u4 u5
RGB source
12 / 40
Laplacians = structure descriptors
X u2 u3 u4 u5
RGB source
Y v2 v3 v4 v5
Luma (‘bad’ color conversion)
13 / 40
Laplacians = structure descriptors
X u2 u3 u4 u5
RGB source
Y v2 v3 v4 v5
Luma (‘bad’ color conversion)
Z t2 t3 t4 t5
‘Good’ color conversion
14 / 40
Laplacians = structure descriptors
X u2 u3 u4 u5 Clustering
RGB source
Y v2 v3 v4 v5 Clustering
Luma (‘bad’ color conversion)
Z t2 t3 t4 t5 Clustering
‘Good’ color conversion
15 / 40
Finding joint eigenbases
Joint approximate diagonalization
Find joint approximateeigenbasis U
minU
off(UTLXU) + off(U
TLYU)
s.t. UTU = I
where off(A) =∑
i 6=j a2ij .
These two problems are equivalent!(approx. joint diagonalizability ⇐⇒ approx. commutativity)
Cardoso 1995, Eynard et al. 2012, Kovnatsky et al. 2013
, Bronstein et al. 2013
16 / 40
Finding joint eigenbases
Joint approximate diagonalization
Find joint approximateeigenbasis U
minU
off(UTLXU) + off(U
TLYU)
s.t. UTU = I
where off(A) =∑
i 6=j a2ij .
Closest commuting Laplacians
Find closest commutingpair LX, LY
minLX,LY
‖LX − LX‖2F + ‖LY − LY‖2F
s.t. LXLY = LYLX
Since LX and LY commute, theyhave a joint eigenbasis U
These two problems are equivalent!(approx. joint diagonalizability ⇐⇒ approx. commutativity)
Cardoso 1995, Eynard et al. 2012, Kovnatsky et al. 2013, Bronstein et al. 201317 / 40
Finding joint eigenbases
Joint approximate diagonalization
Find joint approximateeigenbasis U
minU
off(UTLXU) + off(U
TLYU)
s.t. UTU = I
where off(A) =∑
i 6=j a2ij .
Closest commuting Laplacians
Find closest commutingpair LX, LY
minLX,LY
‖LX − LX‖2F + ‖LY − LY‖2F
s.t. LXLY = LYLX
Since LX and LY commute, theyhave a joint eigenbasis U
These two problems are equivalent!(approx. joint diagonalizability ⇐⇒ approx. commutativity)
Cardoso 1995, Eynard et al. 2012, Kovnatsky et al. 2013, Bronstein et al. 201318 / 40
Laplacian colormaps
X
19 / 40
Laplacian colormaps
X
−→Φθ
Y = Φθ(X)
Parametric colormap Φθ : RNM×d → RNM×d′ parametrizedby θ = (θ1, . . . , θn)
Global: each pixel x is transformed same way, y = Φθ(x)Local: different transformations in q regions,Φθ(X) =
∑qi=1 wiΦθi(X)
20 / 40
Laplacian colormaps
X
−→Φθ
Y = Φθ(X)
LX LY
LX = DX −WX LΦθ(X) = DΦθ(X) −WΦθ(X)
Find an optimal parametric color transformation
minθ∈Rn
‖LXLΦθ(X) − LΦθ(X)LX‖2F + regularization on θ
21 / 40
Color-to-gray conversion
Color mapping by a global color transformation of the form
Φθ(R,G,B) = θ1 + θ2Rθ3 + θ4G
θ5 + θ6Bθ7
Luma Col2Gray Rasche Decolorize Neumann Smith Lu Ours
2.18/-1.05 1.96/-0.10 1.43/-1.38 1.35/0.86 2.22/0.29 2.13/-0.29 1.47/0.82 1.19/1.15
Cadık 2008; Gooch et al. 2005; Rasche et al. 2005; Grundland, Dodgson 2007;
Neumann et al. 2007; Smith et al. 2008; Lu et al. 2012; Kuhn et al. 200822 / 40
Color-to-gray conversion
Color mapping by a global color transformation of the form
Φθ(R,G,B) = θ1 + θ2Rθ3 + θ4G
θ5 + θ6Bθ7
Luma Col2Gray Rasche Decolorize Neumann Smith Lu Ours
5.01/-0.55 3.42/-0.89 3.59/-0.48 3.44/1.41 5.44/-0.66 5.04/-0.19 2.90/0.50 1.28/0.86
9.27/-0.57 7.05/-0.53 7.20/-0.04 7.28/1.45 10.17/-1.05 9.13/-1.02 6.30/1.01 3.78/0.76
Cadık 2008; Gooch et al. 2005; Rasche et al. 2005; Grundland, Dodgson 2007;
Neumann et al. 2007; Smith et al. 2008; Lu et al. 2012; Kuhn et al. 200823 / 40
Color-to-gray conversion
Color mapping by a global color transformation of the form
Φθ(R,G,B) = θ1 + θ2Rθ3 + θ4G
θ5 + θ6Bθ7
Luma Col2Gray Rasche Decolorize Neumann Smith Lu Ours
0.97/0.27 1.24/-1.30 0.97/-0.08 1.02/0.61 1.66/-0.86 1.05/0.32 0.80/0.22 0.85/0.82
Cadık 2008; Gooch et al. 2005; Rasche et al. 2005; Grundland, Dodgson 2007;
Neumann et al. 2007; Smith et al. 2008; Lu et al. 2012; Kuhn et al. 200824 / 40
Color-to-gray conversion
Color mapping by a global color transformation of the form
Φθ(R,G,B) = θ1 + θ2Rθ3 + θ4G
θ5 + θ6Bθ7
Luma Col2Gray Rasche Decolorize Neumann Smith Lu OursRWMS 2.84 2.31 2.46 2.20 4.85 2.94 1.90 1.33z-score -0.17 -0.31 -0.63 0.55 -0.53 -0.09 0.34 0.84
Cadık 2008; Gooch et al. 2005; Rasche et al. 2005; Grundland, Dodgson 2007;
Neumann et al. 2007; Smith et al. 2008; Lu et al. 2012; Kuhn et al. 200825 / 40
Computational complexity: Color-to-gray example
10-1
100
101
102
103
Tim
e (sec)
#vertices 253 641 1130 22946 91784 367136
0.597
RW
MS e
rror
0.599
x10-3
Linear (n=3)
Non-linear (n=7)
Superpixels Scaling
Complexity O(K2)
Laplacian dimension K �MN (realtime performance withsmall K)
Optimization on θ is performed with small Laplacians. Then,Φθ is applied on full image
Superpixels: Ren, Malik 200326 / 40
Color-blind image optimization
RGB source
X
Ψ
Seen by color-blind
Ψ(X)
Vienot et al. 1999, Kim et al. 201227 / 40
Color-blind image optimization
RGB source
X Φθ(X)
Ψ
Seen by color-blind
Φθ
(Φθ ◦Ψ)(X)
︸ ︷︷ ︸‖LXLΦθ(X)−LΦθ(X)LX‖
‖LXL(Φθ◦Ψ)(X)−L(Φθ◦Ψ)(X)LX‖︷ ︸︸ ︷
28 / 40
Color-blind image optimization
RGB source
X Φθ(X)
Ψ
Seen by color-blind
Φθ
(Φθ ◦Ψ)(X)︸ ︷︷ ︸‖LXLΦθ(X)−LΦθ(X)LX‖
‖LXL(Φθ◦Ψ)(X)−L(Φθ◦Ψ)(X)LX‖︷ ︸︸ ︷
29 / 40
Color-blind image optimization: protanopia
RGB Lau
1.23
Optimized
0.50
Lau et al. 201130 / 40
Color-blind image optimization: tritanopia
RGB Lau
1.69
Optimized
0.53
Lau et al. 201131 / 40
Gamut mapping
Map image colors to a gamut G(convex polytope)
minθ∈Rn
‖LXLΦθ(X) − LΦθ(X)LX‖2F+ regularization on θ
s.t. Φθ(X) ⊆ G
sRGB
G
32 / 40
Gamut mapping
Original Lau et al. Ours HPMINDE (clip)
Lau et al. 201133 / 40
RGB+NIR fusion
NIR RGB Lau et al. Ours
Lau et al. 201134 / 40
Multiple image fusion
Morning
Day
Evening
Night
Fusion
35 / 40
Summary
Framework
theoretically grounded
versatile
global/local
realtime
Applications
color-to-grayscale
color-blind optimization
gamut mapping
multispectral image fusion
36 / 40
Summary
Framework
theoretically grounded
versatile
global/local
realtime
Applications
color-to-grayscale
color-blind optimization
gamut mapping
multispectral image fusion
37 / 40
Thank you!
38 / 40
Qualitative evaluation
Web survey
124 volunteers, 2884 pairwise evaluations
Thurstone’s law of comparative judgements → z-score
Consistent with Cadık’s results
39 / 40
Extension: local colormap
RGB Luma Lau et al.
Global Local Clusters
Φθ(X) =∑q
i=1 wiΦθi(X)
Lau et al. 201140 / 40