Gilad Freedman Raanan Fattal Hebrew University of Jerusalem Background and o verview Algorithm description L ocal self similarity Nondyadic filter bank Filter design Results Single ID: 271935
Download Presentation The PPT/PDF document "Image and Video Upscaling from Local Sel..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Image and Video Upscaling from Local Self Examples
Gilad Freedman
Raanan Fattal
Hebrew University of JerusalemSlide2
Background and
o
verview
Algorithm
description
L
ocal
self similarity
Non-dyadic filter bank
Filter
design
ResultsSlide3
Single image upscaling
1.
l
arge
2. realistic
3. faithful4. fastSlide4
Previous workparametric image modelexample based
Freeman et al. 2002
generic looking edges
Sun et al. 2008
Shan et al 2008
Fattal 2007
Glasner
et al. 2009
noisy
resultSlide5
New approach: locale example based
c
orner step edge
l
ine step edge
non
s
mooth shading
local self
s
imilarity
small upscaling ratios
1/2
4/5
new non-dyadic filter bank
Increase exemplar quality and size
maintain search locality
novel components:
l
ocal self similarity
n
on-dyadic filter bankSlide6
Background and overview
Algorithm description
L
ocal
self similarity
Non-dyadic filter bank
Filter
design
ResultsSlide7
Local self-examples upscaling
l
ow
p
ass
original
i
mage
h
igh
p
ass
i
nterpolated image
f
requency
contentSlide8
l
ow pass
h
igh
p
ass
i
nterpolated
i
mage
For each patch:
Search a
local
area for best example
Take corresponding patch
Add to interpolated image
frequency
content
Local
self-examples upscalingSlide9
Local self-examples upscaling
l
ow
p
ass
high pass
i
nterpolated
i
mage
Repeat for all patches, to fill the high frequencies
f
requency
contentSlide10
Overview
Algorithm description
Local self similarity
Non-dyadic filter bank
Filter
design
ResultsSlide11
Local self similaritycropped downscaled
Slide12
Local self similarityPatches in original image can matched locally with ones in downscaled versionSlide13
Local examples are enoughfull imageimage
d
atabase
query
db
i
mage
l
ocal
4.0 2.9 3.55
1.6 1.05 1.05
2.7 2.05 2.05
3.3 2.96 3.06
6.5 5.61 5.61
best
m
atchesSlide14
Visual assessment – external, exact NN, localLarge externalexample database
Searching the
entire image
Searching local
regions in image
e
xternal database
g
lobal search
local
s
earchSlide15
Comparison of example search methodsSlide16
Background and overview
Algorithm description
L
ocal
self similarity
Non-dyadic
filter bank
Filter
design
ResultsSlide17
Need for non-dyadic scalings
large ratios
mixed ratios
small ratiosSlide18
Dyadic filters
1:2
full frequency content
h
igher half
l
ower half
dyadic filter bankSlide19
Non-dyadic filter bank
1:2
4:5
f
ull frequency content
h
igher part
l
ower part
n
on-dyadic filter bankSlide20
Non-dyadic filters: downscaling
1. convolve with 2 filters
2. subsample each by 3
d
yadic case:
e
xample for the 2:3 ratio:
1. convolve with one filter
2. subsample by 2Slide21
Non-dyadic filters: upscaling
1. zero upsample by 2
2
. convolve with
2 filters
3. sum
d
yadic case:
e
xample for the 2:3 ratio:
1. zero upsample by
1
2. convolve with 1 filterSlide22
Use of the filters in upscaling
Upscaling using inverse scaling filters
Smoothing by downscaling and upscalingSlide23
Background and overview
Algorithm description
L
ocal
self similarity
Non-dyadic filter bank
Filter design
ResultsSlide24
When interpolating, smooth areas come from
input
Uniformly
spaced
grids should remain
uniform
1. Uniform stretch
0
255
brightness
grid coordinatesSlide25
2. Consistency
upsample
d
ownsample
The interpolated image, if downscaled should be equal to the input.
Formally,
Previous methods achieve consistency by solving large linear systems to
achieve this propertySlide26
3. PSF modeling
L
arge image - small camera point spread function
Small image - large camera point spread function
Difference between
point spread functionsSlide27
4. Low frequency span
frequency
When upsampling don’t add new frequencies
Upsampling
filter should be low-pass
o
riginal
interpolatedSlide28
5. Singularities preservation
blurred Image
interpolated image
s
imilar
amount of blurSlide29
Real time video upsampling on GPU
m
ain GPU memory
GPU cores
NTSC to full HD @ 24 fps
Search and filter-banks
are both
local
operationsSlide30
Background and overview
Algorithm description
L
ocal
self similarity
Non-dyadic filter bank
Filter
design
ResultsSlide31
Bicubic
x3
(zoomed in
)
Ours X3
(zoomed in)Slide32
Bicubic
x3
Ours X3Slide33
Genuine Fractals™ x4
Ours X4Slide34
Glasner
et al. 2009 x4
Ours X4Slide35Slide36Slide37
Thank you!Paper & additional results can be found at: www.cs.huji.ac.il/~giladfreedmn