Upload
hojjat-seyed-mousavi
View
70
Download
4
Embed Size (px)
Citation preview
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Sparsity Based Super ResolutionUsing Color Channel Constraints
Hojjat Mousavi, Vishal Monga
School of Electrical Engineering and Computer Science
The Pennsylvania State University
September 27, 2016
iPAL Color Super Resolution 1/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Super Resolution - Problem Definition
• Multi-frame SR: Traditional Super-resolution problem is the process ofcombining multiple low resolution images to form a higher resolution one
• Resulting image should represent reality better than all the input images.
• Single-image SR: given a single low-resolution input, reconstruct ahigh-resolution version of the input.
• Advantage: more widely applicable than multi-frame approaches.• Challenge: single-image super-resolution is an extremely ill-posed
problem.
iPAL Color Super Resolution 1/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Sparsity-based Super resolution - Basic idea 1,2
• Construct two coupled dictionaries based on image patches inluminance (Y) channel
1 Low resolution dictionary: DDDl (High frequency features)
2 High resolution dictionary: DDDh (Actual high resolution patches)
• Atoms of each dictionary correspond to each other and are LR-HRcounterparts of each other extracted from the same locations
1Wright et al. CVPR 20082Wright et al. TIP 2009
iPAL Color Super Resolution 2/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Sparsity-based Super resolution
• SR for an unseen low resolution image:
1 Extract overlapping patches (overlapped tiling) (yyyl)
2 For each patch find the low resolution representation using DDDl
xxx∗ = argminxxx
12||yyyl−DDDlxxx||22 +λ ||xxx||1
Find the sparse linear representation of low resolution patch based on LRdictionary
3 Find the high resolution representation using DDDh and the same xxx∗.
yyyh =DDDhxxx∗
4 construct the high resolution image from high-res patches.
iPAL Color Super Resolution 3/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Color Super Resolution
• YCbCr space. Apply Bicubic interpolation on Cb and Cr channels.• Human eye is more sensitive to luminance than chrominance• Some images have varying amount of luminance and chrominance
geometry
• Chrominance channels also contain useful information• Super-resolution only on luminance channel may not get the best results• Luminance edge (in Y)→ present in R, G and B channels• Jointly account for cross channel information in an adaptive manner
iPAL Color Super Resolution 4/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Color Super Resolution
• How to capture edge similarities?• Extract edge information in RGB channels
• Find patches that should have have high edge correlation based on amountof color information in each patch
• Encourage edge similarity in selected patches of high resolution image
• Similarity (Correlation) between edges in different channels. (Based on HRimage): Example: ‖SSSryyyhr −SSSgyyyhg‖2 Or correlation (SSSryyyhr )
T(SSSgyyyhg)
where SSSr,SSSg,SSSb are highpass edge detector filters
• Color constraints: Edge differences across color channels are minimizedfor selected patches3,4,5
‖SSSryyyhr −SSSgyyyhg‖2 < εrg
‖SSSgyyyhg −SSSbyyyhb‖2 < εgb
‖SSSbyyyhb −SSSryyyhr‖2 < εbr3Srinivas et al. CIC 20104Farsiu et al. TIP 20065Menon et al. TIP 2009
iPAL Color Super Resolution 5/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Color Super Resolution
• High resolution representation of patches is incorporated in the costfunction by the following assumption of conventional SR methods.
yyyhr =DDDhrxxxr, yyyhg =DDDhgxxxg, yyyhb =DDDhbxxxb
• Incorporating RGB channel information and exploiting our multi-taskframework result in the following optimization problem:
argminxxxc
∑c={r,g,b}
(12‖yyylc −DDDlcxxxc‖2
2 +λ‖xxxc‖1
)+τ
[‖SSSrDDDhrxxxr−SSSgDDDhgxxxg‖2
2 +‖SSSgDDDhgxxxg−SSSbDDDhbxxxb‖22 +‖SSSbDDDhbxxxb−SSSrDDDhrxxxr‖2
2
].
• Note: Without color channel constraints→ Three independentoptimization problems
• With additional color constraints→ One optimization problem withquadratic constraints on pairs of channels
• τ is very crucial and we pick it in an adaptive manneriPAL Color Super Resolution 6/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Color Super Resolution
• Introducing DDD,DDDl,xxx,yyy, we can simply the cost function to
xxx = argminxxx xxxTDDDxxx−yyyTDDDlxxx+λ‖xxx‖1. → FISTA6
where
DDD =
12 DDDT
lrDDDlr +2τDDDT
hrSSST
r SSSrDDDhr −2τDDDThr
SSSTr SSSgDDDhg 000
000 12 DDDT
lgDDDlg +2τDDDT
hgSSST
g SSSgDDDhg −2τDDDThg
SSSTg SSSbDDDhb
−2τDDDThb
SSSTb SSSrDDDhr 000 1
2 DDDTlb
DDDlb+2τDDDT
hbSSST
b SSSbDDDhb
xxx =
xxxrxxxgxxxb
, yyy =
yyylryyylgyyylb
, DDDl =
DDDlr 000 000000 DDDlg 000000 000 DDDlb
• Note that matrix DDD can capture cross channel constraints by adding a
term to the appropriate locations• SSSr,SSSg,SSSb are gradient operators in RGB channels.6Beck et al. SIAM Journal of Imaging Sciences, 2009
iPAL Color Super Resolution 7/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Joint Dictionary Learning for Color Channels
• Correlation between color channels can be even better captured if theindividual color channel dictionaries are also designed to facilitate thesame.
• Given a set of N sampled training image patch pairs {YYYh,YYY l}.YYYh = {yyy1
h,yyy2h, ...,yyy
Nh }: set of HR patches sampled from training
YYY l = {yyy1l ,yyy
2l , ...,yyy
Nl }: set of corresponding LR patches.
• A new joint learning for multi channel dictionary learning is proposed:
arg minDDDh,DDDl,xxxi
1N
N∑i=1
γ
2‖yyyi
l−DDDlxxxi‖22 +
1− γ
2‖yyyi
h−DDDhxxxi‖22
+τ
[‖SSSrDDDhrxxx
ir−SSSgDDDhgxxxi
g‖22
+‖SSSgDDDhgxxxig−SSSbDDDhbxxxi
b‖22
+‖SSSbDDDhbxxxib−SSSrDDDhrxxx
ir‖2
2
]+λ‖xxxi‖1
st. ‖DDDh(:,k)‖22 ≤ 1, ‖DDDl(:,k)‖2
2 ≤ 1, k = 1,2, ...,K
iPAL Color Super Resolution 8/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Joint Dictionary Learning for Color Channels
L2 =1N
N∑i=1
γ
2‖yyyi
l−DDDlxxxi‖22 +
1− γ
2‖yyyi
h−DDDhxxxi‖22 +λ‖xxxi‖1
+2τxxxiT DDDThSSST(III−PPPT
s )SSSDDDhxxxiT
=γ
2N‖YYY l−DDDlXXX‖2
F +1− γ
2N‖YYYh−DDDhXXX‖2
F +λ
N‖XXX‖1
+2τ
NTr(
XXXTDDDThSSST(III−PPPT
s )SSSDDDhXXXT).
where XXX = [xxx1 xxx2 ... xxxN ].Alternatively solve for XXX, DDDl and DDDh
xxx =
xxxrxxxgxxxb
, yyy =
yyylryyylgyyylb
, DDDl =
DDDlr 000 000000 DDDlg 000000 000 DDDlb
, DDDh =
DDDhr 000 000000 DDDhg 000000 000 DDDhb
SSS =
SSSr 000 000000 SSSg 000000 000 SSSb
, PPPs =
000 000 IIIp2×p2
IIIp2×p2 000 000000 IIIp2×p2 000
iPAL Color Super Resolution 9/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Solution for XXX
• DDDl and DDDh fixed.• Optimize over XXX whose columns can be obtained independently.• For each column of XXX (i = 1...N) we can rewrite the cost function as:
xxxi = argminxxx
γ
2‖yyyi
l−DDDlxxx‖2F +
1− γ
2‖yyyi
h−DDDhxxx‖2F +λ‖xxx‖1
+2τxxxTDDDThSSST(III−PPPT
s )SSSDDDhxxxT
= argminxxx
xxxT [γ
2DDDT
l DDDl +1− γ
2DDDT
hDDDh
+2τDDDThSSST(III−PPPT
s )SSSDDDh]xxx
−(γyyyiT
l DDDl +(1− γ)yyyiTh DDDh
)xxx+λ‖xxx‖1
= argminxxx
xxxTAAAxxx−bbbTxxx +λ‖xxx‖1
AAA = γ
2DDDTl DDDl +
1−γ
2 DDDThDDDh +2τDDDT
hSSST(III−PPPTs )SSSDDDh
bbbiT = γyyyiTl DDDl +(1− γ)yyyiT
h DDDh.• Can be solved using FISTA7
7Beck et al. SIAM Journal of Imaging Sciences, 2009
iPAL Color Super Resolution 10/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Solution for DDDl
• Fix XXX and DDDh, the cost function reduces to:
DDDl = argminDDDl
‖YYY l−DDDlXXX‖2F
s.t. ‖DDDl(:,k)‖22 ≤ 1, k = 1,2, ...,K
DDDl is block diagonal .
• Split into three separate dictionary learning procedures as follows wherec ∈ {r,g,b}.
DDDlc = argminDDDlc
‖YYY lc −DDDlcXXXc‖2F
s.t. ‖DDDlc(:,k)‖22 ≤ 1, k = 1,2, ...,K
where XXXc = [xxx1c xxx2
c ... xxxNc ], YYY lc = [yyy1
c yyy2c ... yyyN
c ] and c takes the subscripts from{r,g,b} indicating a specific color channel.
• Each of the above dictionaries are learnt by the ODL method8.8Mairal et al. ICML, 2009.
iPAL Color Super Resolution 11/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Solution for DDDh• Finally when XXX and DDDl are fixed, L2 reduces to:
DDDh = argminDDDh
1N
N∑i=1
1− γ
2‖yyyi
h−DDDhxxxi‖22
+2τxxxiT DDDThSSST(III−PPPT
s )SSSDDDhxxxiT
s.t ‖DDDh(:,k)‖22 ≤ 1, k = 1,2, ...,K.
• Not very straight forward to solve→ ADMM9.• Define the function g(DDDh,ZZZ) as follows:
g(DDDh,ZZZ) =1N
N∑i=1
1− γ
2‖yyyi
h−DDDhxxxi‖22 +2τxxxiT DDDT
hSSST(III−PPPTs )SSSZZZxxxiT
• Solve the equivalent bi-convex problem:DDDh = argmin
DDDh,ZZZg(DDDh,ZZZ)
s.t DDDh−ZZZ = 000,‖DDDh(:,k)‖2
2 ≤ 1, k = 1,2, ...,K.9Boyd et al. Foundations and Trends in Machine Learning, 2011
iPAL Color Super Resolution 12/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Solution for DDDh
Iterative steps of ADMM until a convergence is achieved are as follows:1 Find DDDt+1
h :DDDt+1
h = argminDDDh
( 1N
N∑i=1
1− γ
2‖yyyi
h−DDDhxxxi‖22
+2τxxxiT DDDThSSST(III−PPPT
s )SSSZZZtxxxiT)
+ρ
2‖DDDh−ZZZt +UUUt‖2
F
s.t. ‖DDDh(:,k)‖22 ≤ 1, k = 1, ...,K.
2 Find ZZZt+1:ZZZt+1 = argmin
ZZZ
(2τ
N
N∑i=1
xxxiT DDDt+1T
h SSST(III−PPPTs )SSSZZZtxxxiT
)+
ρ
2‖DDDt+1
h −ZZZ +UUUt‖2F
3 Find UUUt+1: UUUt+1 =UUUt +DDDt+1h −ZZZt+1
Solutions to steps 1 and 2 of the ADMM procedure are not straightforwardand details are in the paper.
iPAL Color Super Resolution 13/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
• More analytical Results on how to solve optimization problems at eachstage
• Extensive experimental validations in addition to high quality images
• Implementation and MATLAB toolbox
All Available online at:http://signal.ee.psu.edu/MCcSR.html
iPAL Color Super Resolution 14/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
State of the art methods to compare with
Single image super resolution methods that incorporate sparsity methods:1 Sparsity Constrained super resolution (ScSR)10
2 Single Image Scale-up using Sparse Representation 11
3 Adjusted Anchored Neighborhood Regression for Fast Example-BasedSuper-Resolution (ANR+)12
4 Global Regression for Fast Super-Resolution (GR)13
5 Neighbor Embedding with Locally Linear Embedding (NE+LLE) 14
6 Neighbor Embedding with NonNegative Least Squares (NE+NNLS) 15
7 Single Image SR using sparse regression and natural image prior16:Using sparse kernel ridge regression and natural image priors.
8 Image and Video Upscaling from Local Self-Examples17
10Yang, IEEE TIP, 201211Zeyde et al, Springer, Curves and Surfaces, 201212Timofte et al. ACCV 201413Timofte et al. ICCV 201314Chang et al. CVPR 200415Bevilazqua et al. BMVC 201216Kim et al. IEEE Tran on PAMI, 201017Freeman et al, ACM Transactions on Graphics, 2011
iPAL Color Super Resolution 15/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Experimental Set Up
• Dictionary size: 512 atoms - 100,000 image patches are sampled• Scaling factor: 2x, 3x, 4x• Patch size: 5×5, 7×7, 9×9 pixels.• Quantitative measures: PSNR, SSIM, S-CIELAB18
18Zhang et al., in Proc. IEEE COMPCON Symp. Dig., 1997.
iPAL Color Super Resolution 16/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
(a) PSNR and SSIM - scale 2:Top: Original, Bicubic (30.46, 0.840), Zeyde et al. (31.97, 0.887),Middle: GR (31.70, 0.879), ANR (32.09, 0.889), NENNLS (31.87, 0.884)Bottom: NELLE (32.03, 0.889), MCcSR (32.23, 0.899), ScSR (32.14,0.893) .
(b) SCIELAB error map - scale 2:Top: Original, Bicubic (1.898e4), Zeyde et al. (1.127e4),Middle: GR (1.198e4), ANR (1.077e4), NENNLS (1.159e4)Bottom: NELLE (1.099e4), MCcSR (9.770e3) , ScSR (1.014e4).
iPAL Color Super Resolution 17/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
(c) PSNR and SSIM - scale 3:Top: Original, Bicubic (27.51, 0.685), Zeyde et al. (28.28, 0.737),Middle: GR (28.15, 0.729), ANR (28.36, 0.742), NENNLS (28.17, 0.730)Bottom: NELLE (28.30, 0.738), MCcSR (28.51, 0.758), ScSR (28.25,0.740) .
(d) SCIELAB error map - scale 3:Top: Original, Bicubic (3.423e4), Zeyde et al. (2.896e4),Middle: GR (3.008e4), ANR (2.865e4), NENNLS (2.961e4)Bottom: NELLE (2.905e4), MCcSR (2.709e4) , ScSR (3.002e4).
iPAL Color Super Resolution 18/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
(e) PSNR and SSIM - scale 4:Top: Original, Bicubic (26.05, 0.566), Zeyde et al. (26.61, 0.615),Middle: GR (26.51, 0.607), ANR (26.63, 0.618), NENNLS (26.50, 0.606)Bottom: NELLE (26.57, 0.614), MCcSR (26.74, 0.632), ScSR (26.35,0.608) .
(f) SCIELAB error map - scale 4:Top: Original, Bicubic (4.369e4), Zeyde et al. (3.923e4),Middle: GR (4.045e4), ANR (3.928e4), NENNLS (3.984e4)Bottom: NELLE (3.967e4), MCcSR (3.818e4) , ScSR (4.002e4).
iPAL Color Super Resolution 19/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Results - Scale 3
Table: PSNR results of different methods for various images with scaling factor of 3.
Images PSNR (dB)Bicub Zeyde GR ANR NENNLS NELLE MCcSR ScSR
baby 38.42 39.51 39.38 39.56 39.22 39.49 39.51 39.40butterfly 28.73 30.60 29.73 30.57 30.29 30.42 30.59 30.64bird 36.37 37.90 37.44 37.92 37.68 37.90 38.02 37.59face 35.96 36.44 36.40 36.50 36.39 36.47 36.48 36.37foreman 35.76 37.67 36.84 37.71 37.37 37.69 37.74 37.64coastguard 31.31 31.91 31.78 31.84 31.77 31.83 31.95 31.83flowers 30.92 31.84 31.62 31.88 31.68 31.80 32.07 31.87head 36.02 36.47 36.42 36.52 36.40 36.50 36.51 36.42lenna 35.26 36.23 35.99 36.29 36.11 36.24 36.33 36.14man 31.78 32.68 32.44 32.71 32.50 32.65 32.75 32.68pepper 35.25 36.27 35.77 36.13 35.99 36.12 36.30 36.20average 33.08 34.06 33.76 34.07 33.88 34.03 34.14 34.00
iPAL Color Super Resolution 20/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Results - Scale 3
Table: SSIM results of different methods for various images with scaling factor of 3.
Images SSIMBicub Zeyde GR ANR NENNLS NELLE MCcSR ScSR
baby 0.88 0.90 0.90 0.90 0.89 0.90 0.90 0.89butterfly 0.79 0.85 0.80 0.84 0.84 0.84 0.85 0.85bird 0.90 0.92 0.91 0.92 0.92 0.92 0.93 0.91face 0.72 0.74 0.74 0.74 0.74 0.74 0.75 0.74foreman 0.89 0.91 0.90 0.91 0.90 0.91 0.91 0.90coastguard 0.57 0.62 0.63 0.62 0.61 0.62 0.63 0.62flowers 0.77 0.80 0.79 0.80 0.79 0.80 0.81 0.80head 0.72 0.74 0.74 0.75 0.74 0.74 0.75 0.74lenna 0.78 0.80 0.80 0.80 0.80 0.80 0.81 0.80man 0.72 0.76 0.76 0.77 0.76 0.76 0.76 0.76pepper 0.78 0.80 0.79 0.80 0.79 0.79 0.80 0.79average 0.745 0.776 0.769 0.778 0.771 0.775 0.785 0.774
iPAL Color Super Resolution 21/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Results - Scale 3
Table: S-CIELAB error results of different methods for various images with scalingfactor of 3.
Images S-CIELABBicub Zeyde GR ANR NENNLS NELLE MCcSR ScSR
baby 2.07E+04 1.36E+04 1.40E+04 1.32E+04 1.47E+04 1.34E+04 1.34E+04 1.50E+04butterfly 2.28E+04 1.55E+04 1.84E+04 1.55E+04 1.60E+04 1.60E+04 1.54E+04 1.49E+04bird 1.07E+04 7.36E+03 8.02E+03 7.21E+03 7.73E+03 7.30E+03 6.50E+03 7.81E+03face 3.79E+03 2.71E+03 2.73E+03 2.57E+03 2.73E+03 2.61E+03 2.47E+03 2.70E+03foreman 8.46E+03 3.90E+03 4.79E+03 3.48E+03 4.01E+03 3.62E+03 3.72E+03 3.89E+03coastguard 1.96E+04 1.71E+04 1.70E+04 1.70E+04 1.76E+04 1.71E+04 1.69E+04 1.70E+04flowers 4.47E+04 3.75E+04 3.89E+04 3.69E+04 3.84E+04 3.74E+04 3.29E+04 3.70E+04head 3.79E+03 2.69E+03 2.74E+03 2.54E+03 2.79E+03 2.61E+03 2.42E+03 2.65E+03lenna 2.44E+04 1.74E+04 1.85E+04 1.67E+04 1.79E+04 1.69E+04 1.58E+04 1.72E+04man 3.80E+04 2.91E+04 3.03E+04 2.84E+04 3.02E+04 2.89E+04 2.88E+04 2.95E+04pepper 2.48E+04 1.91E+04 2.15E+04 1.96E+04 2.02E+04 1.95E+04 1.73E+04 1.91E+04
average 2.79E+04 2.27E+04 2.36E+04 2.24E+04 2.33E+04 2.26E+04 2.14E+04 2.28E+04
iPAL Color Super Resolution 22/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Effect of RGB constraints
Figure: Visual Images as well as S-CIELAB error maps are shown for a scaling factorof 3. From left to right for each row images correspond to: Original image, applyingSR separately on RGB channels (36.26, 0.83, 1.57e4), ScSR (36.13, 0.83, 1.67e4)and MCcSR (36.67, 0.85, 1.43e4). Numbers in parenthesis are PSNR, SSIM andSCIELAB error measures.
iPAL Color Super Resolution 23/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Effect of RGB constraints
Figure: Visual Images as wellas S-CIELAB error maps areshown for a scaling factor of 3.From left to right for each rowImages correspond to: OriginalImage, applying SR separatelyon RGB channels, ScSR,MCcSR
Table: Quantitative measures to show effectiveness of color constraints in SR for a scaling factor of 3.
Images PSNR SSIM S-CIELABSeparate RGB ScSR MCcSR Separate RGB ScSR MCcSR Separate RGB ScSR MCcSR
comic 28.37 28.25 28.51 0.74 0.74 0.75 2.80e4 3.00e4 2.70e4baboon 26.95 26.95 27.11 0.53 0.52 0.54 9.93e4 1.01e5 9.57e4pepper 36.14 35.85 36.30 0.79 0.77 0.81 1.93e4 2.27e5 1.73e4bird 37.71 37.59 38.02 0.92 0.912 0.93 7.28e3 8.54e3 6.50e3
iPAL Color Super Resolution 24/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Performance Under Noise
Figure: Super Resolution performance under difference noisestandard deviations: 4,6,8,12 (from top to bottom ) with different methods:Original, bicubic, MCcSR, ScSR (from left to right)
Table: Average performance under differentnoise levels.
Measure Method σ = 0 σ = 4 σ = 6 σ = 8 σ = 12
PSNRBicubic 33.08 32.99 32.75 32.50 31.88ScSR 34.00 33.95 33.92 33.90 33.86
MCcSR 34.14 34.11 34.09 34.09 34.07
SSIMBicubic 0.745 0.731 0.698 0.672 0.619ScSR 0.774 0.772 0.766 0.761 0.752
MCcSR 0.785 0.783 0.780 0.775 0.768
SCIELABBicubic 2.79E4 2.92E4 4.40E4 5.25E4 6.31E4ScSR 2.28E4 2.31E4 2.36E4 2.39E4 2.43E4
MCcSR 2.14E4 2.16E4 2.20E4 2.21E4 2.23E4
iPAL Color Super Resolution 25/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Dictionary size
• Dictionaries of size 16,32,64,128,256 and 512 are trained.
Figure: Effect of dictionary size on PSNR, SSIM and S-CIELAB erroriPAL Color Super Resolution 26/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Conclusion and Future Work
• Sparsity→ powerful prior for the ill-posed problem of single image superresolution
• Cross channel information and color constraints→ Regularizing theoptimization problem for boosting SR performance
• Under different scaling factors, different noise levels, different dictionarysizes the proposed MCcSR method outperforms the state of the art.
• Expedite the sparse coding problem using neural networks
• Introduce other objective measurements rather than MSE for qualityassessment or in the objective function
• Apply other cross channel constraints or color constraint that canimprove super resolution performance
iPAL Color Super Resolution 27/28
Background Color SR Color Dictionary Learning Experimental Results Conclusion
Thank you!
iPAL Color Super Resolution 28/28
Backup Slides
iPAL Color Super Resolution 1/6
Dictionary Atoms
• Low resolution dictionary atoms - Red channel only- 9 by 9 pixelspatches - Extracted from features
Figure: Low resolution dictionary atoms -Red channel only- 9 by 9 pixels patches -Extracted from features
Figure: High resolution dictionary atoms -RGB - 9 by 9 pixels patches
iPAL Color Super Resolution 2/6
Color Adaptive Patch Processing
• Not all image patches have the same color information (chrominance)
• Parameter τ can be used to control high frequency correlation amongcolor channels.
• Calculate the variations in color information→ Adaptively control τ.
β =12s
(‖HHH1yyyCb‖+‖HHH1yyyCr‖‖HHH1yyyY‖
+‖HHH2yyyCb‖+‖HHH2yyyCr‖
‖HHH2yyyY‖
)where s is normalization parameter, HHH1 and HHH2 are high-pass Scharroperators.
iPAL Color Super Resolution 3/6
(a) PSNR and SSIM - scale 2:Top: Original, Bicubic (28.19, 0.635), Zeyde et al. (28.62, 0.683),Middle: GR (28.63, 0.690), ANR (28.67, 0.689), NENNLS (28.58, 0.680)Bottom: NELLE (28.66, 0.688), MCcSR (28.78, 0.705, ScSR (28.69, 0.692).
(b) SCIELAB error map - scale 2:Top: Original, Bicubic (7.856e4), Zeyde et al. (6.570e4),Middle: GR (6.388e4), ANR (3.287e4), NENNLS (6.585e4)Bottom: NELLE (6.421e4), MCcSR (5.799e4) , ScSR (6.296e4).
iPAL Color Super Resolution 4/6
(c) PSNR and SSIM - scale 3:Top: Original, Bicubic (26.71, 0.480), Zeyde et al. (26.94, 0.520),Middle: GR (26.95, 0.529), ANR (26.97, 0.527), NENNLS (26.92, 0.518)Bottom: NELLE (26.97, 0.526), MCcSR (27.11, 0.549), ScSR (26.95,0.524) .
(d) SCIELAB error map - scale 3:Top: Original, Bicubic (1.078e5), Zeyde et al. (1.008e5),Middle: GR (1.000e5), ANR (9.962e4), NENNLS (1.010e5)Bottom: NELLE (9.998e4), MCcSR (9.574e4) , ScSR (1.018e5).
iPAL Color Super Resolution 5/6
(e) PSNR and SSIM - scale 4:Top: Original, Bicubic (26.00, 0.390), Zeyde et al. (26.17, 0.420),Middle: GR (26.17, 0.428), ANR (26.19, 0.426), NENNLS (26.15, 0.419)Bottom: NELLE (26.18, 0.425), MCcSR (26.25, 0.446), ScSR (26.11,0.415) .
(f) SCIELAB error map - scale 4:Top: Original, Bicubic (1.237e5), Zeyde et al. (1.186e5),Middle: GR (1.183e5), ANR (1.180e5), NENNLS (1.190e5)Bottom: NELLE (1.183e5), MCcSR (1.136e5) , ScSR (1.185e5).
iPAL Color Super Resolution 6/6