super resolution from a single image
TRANSCRIPT
Super-Resolution from a Single Image
by Daniel Glasner, Shai Bagon and Michal Irani
Presented by : Lakkhana MallikarachchiSupervised by : Prof. N.D. Kodikara
Authors
Daniel Glasner
› B.Sc and M.Sc at Tel Aviv University, Israel.› Reading for a PhD at The Weizmann Institute of
Science, Israel.› Web - http://www.wisdom.weizmann.ac.il/~glasner/
Shai Bagon
› PhD at The Weizmann Institute of Science, Israel.› Web - http://www.wisdom.weizmann.ac.il/~bagon/
Authors
Michal Irani› Professor in Computer Science at The
Weizmann Institute of Science, Israel.› Research Interests - Computer Vision and Video
Information Analysis› Web - http://www.wisdom.weizmann.ac.il/~irani/
The need of Super Resolution Main approaches of SR Proposed method Experimental results of the proposed
method
Outline
Why Super Resolution?
Low Resolution
High Resolution
Resize
Generating high resolution imagewith more resolving powerusing one or more low resolution
images.
› more resolving power – more details
What Is Super Resolution?
Multi image super resolution
Example based super resolution
Super Resolution Methods
Several images of the same scenery.
Each image will have different information of the same scenery.
Multi Image Super Resolution
Image database with HR/LR image pairs
Replace similar LR patches with corresponding HR patches.
Example Based Super Resolution
+
LR HR
Combine multi image SR with example based SR
Without use external source
Proposed Approach
5x5 pixel image patches
More than 60% of image patches have 9 or more recurrences within same scale or in different scales
Patch Redundancy
Use patch redundancy in same scale to model multi image super resolution problem
Use patch redundancy in different scales to model example based super resolution problem
Proposed Method
Problem Model
L1 L2 L3
H
Problem Model
p
q
Find similar patches within scale Nearest Neighbor
Multi Image to Single Image
How to find cross-scale patch redundancy?
Finding Similar Patches
I0 = L
I-1
I1 = H
Finding Similar Patches
I0 = L
I-1
I1
I2 = H
Finding Similar Patches
FindNN
FindNN
𝑝
~𝑝
I0 = L
I-1
I-2
I1
I2 = H
Finding Similar Patches
FindNN P
are
nt
FindNN Parent
𝑝
~𝑝
I0 = L
I-1
I-2
I1
I2 = H
𝑠𝑙 .~𝑝
Finding Similar Patches
FindNN P
are
nt
Copy
FindNN Parent Copy
𝑝
~𝑝
𝑠𝑙 .~𝑝
𝑠𝑙 .𝑝
I0 = L
I-1
I-2
I1
I2 = H
RGB YIQ Extract Y component (Luminance) Apply SR to Y component Use interpolation methods to I and Q
components (Chrominance) Combine YIQ
Color Images
Experimental Results
Bi-cubic interpolation Proposed Method
LR
Experimental Results
Nearest NeighborProposed Method
LR
Experimental Results
Example BasedProposed Method
LR
Experimental Results
Bi-cubic interpolation Proposed Method
LR
Two main approaches of Super Resolution
Observation about patch redundancy Unified approach of Super Resolution Experimental results
Summery
D. Glasner, S. Bagon and M. Irani, "Super-resolution from a single image," in IEEE 12th International Conference on Computer Vision (ICCV 2009), Kyoto, Japan, Sep. 29 - Oct. 2, 2009, pp. 349-356.
http://www.wisdom.weizmann.ac.il/~vision/SingleImageSR.html
http://cs.brown.edu/courses/csci1950-g/results/final/pachecoj
References
Any Question?
Thank you…