Noisy Video Super-Resoluti on
Feng Liu, JinjunWang,ShenghuoZhu (MM’08)
University of Wisconsin-Madison, NEC Laboratories America, Inc.
第一組: 資訊四 B95902105 黃彥達 資訊碩一 R98922046 蔡旻光 網媒碩二 R97944012 鄒志鴻
Outline
Introduction
Goal
File Format
Noise Reduced Image
Proposed Approach
Motion Estimation & Estimated Super-Resol ution Result
Implementation
Result
Conclusion
2
Introduction
Low-quality videos often not only have lim ited resolution but also suffer from noise
In fact, the requirements of de-noising & super- resolution is quite similar
This paper present a unified framework wh ich achieves simultaneous video de-noisin g and super-resolution algorithm by some m easurements of visual quality
3
Goal
Refine low-quality videos from YouTub e, and make the video better effects, which has better quality by human eye s.
Input is low-quality and noise-includ
ed (block effects or somewhat noise)
videos
Noise-Reduced Image
mv-SAD Gaussian- space
Gaussian- time
| p(I,j) – p(i’, j’) | >
threshold
Gaussian Space
Frame t
Pixel(I ,j)
Standard deviation Set Mean
= 0
Motion Vector
Frame t Pixel ( i , j , t)
Frame Pixel ( i + mv_i , j + mv_j t+1 , t+1)
(mv_i , mv_j)
Gaussian Time
Frame t - 2
Frame t - 1
Frame t
Space Gaussian
Time Gaussian
Pixel(I ,j)
Frame t+1
Frame t+2
Frame t
Noise-Reduced Image
Before After
Proposed Approach – 1 / 4
Consider the visual quality with respect t o the following 3 aspects:
Fidelity Preserving
▪ To achieve similar high-resolution result
Detail Preserving
▪ Enhanced details (edge)
Spatial-Temporal Smoothness
▪ Remove undesirable high-frequency contents (e.g. jitter)
10
Proposed Approach – 2 / 4
Fidelity Preserving
Conventional metrics:
▪ Measure fidelity by the difference between Ih & Il would be problematic & waste useful time-space information in video
Proposed metrics:
▪ Estimate an approximation of super-resolution results from space-time neighboring pixels
▪ The fidelity measurement:
see next page for details noised
11
Proposed Approach – 3 / 4
Detail Preserving
Enhanced details (edge)
Contrast preserving
▪ Human visual system is more sensitive to contrast than pixel values
▪ Gradient fields of Ih & should be close
,where Wk is one or zero if the patchk with/o edges (canny detector)
12
Proposed Approach – 4 / 4
(Spatial-Temporal) Smoothness
Smooth results are often favored by the human system
Encourage to minimize:
A 2-D Laplace filter may be
Spatial-temporal Laplacian
OR
13
An Optimization Problem
Proposed Measurements
A quadratic minimization problem to solve (AX = b):
Contrast Similarity
Detail Information(edge)
Spatial-Temporal Smoothness
14
Implementation – 1 / 2
Inputlow
Ih~
Ih~
Gh~
Ih~
) 1 , ( )
1 ,
(I~ t w I~ t
wmv h mv h
X =
I
6 -1 … -1 -1 6 -1 … -1
-1 6 -1 … -1Laplacian
Gradient
-1 0 1 … 1 -1 0 1 … 1 -1 0 1 … 1
Minimize Edge
Motion Estimation Motion Estimation
++
Result (X)
Fidelity
Bilateral filter
wmv
fidv
fidg
dt
sm
15
Implementation – 2 / 2
Adjustments for the weight terms
The measurement term is more emphasized if the w eight is larger
By iteratively experiments for our test data, we took
However, we found that for different videos, the best weight sets may be also different
3 . 0 ,
1 . 0 ,
1 ,
1
fidg dt sm
fidv
16
Result
352 x 288 Result
17
Conclusion
The proposed framework formulates noisy vi deo super-resolution as an optimization pr oblem, aiming to maximize the visual quali ty of the result
The measurements of fidelity-preserving, d etail-preserving and smoothness are consid ered to maximize the visual quality result s
18
Thank you!!
19