An Improved 3DRS Algorithm for
Video De-interlacing
Songnan Li, Jianguo Du, Debin Zhao, Qian Huang, Wen Gao
in IEEE Proc. Picture Coding Symposium (PCS), Beijing, China, Apr. 2006.

INTRODUCTION

Improved 3DRS Algorithm


Bi-directional Block Matching
Multi-resolution Searching

EXPERIMENTAL RESULTS

CONCLUSION
Introduction

Traditional television systems adopted interlaced scan




As a tradeoff between frame rate and transmission bandwidth
requirements
Some artifacts, such as
flicker, serration
makes video coding
less efficient and
more complex
de-interlacing as a picture
format conversion from
interlaced to progressive pictures
De-interlacing methods(1/2)

Line doubling ,
line interpolation



frame rate identical to
the field rate
used more frequently in
high end consumer electronics
Weaving

forming a jagged edge
De-interlacing methods(2/2)

Directional dependent filtering


Adaptive adjust
Motion-Compensated (MC) interpolation



better results
massive requirements on computational resources and storage.
3D recursive searching (3DRS) algorithm has superiority
at both true motion estimation and hardware complexity


forward block matching suffers from latent error propagation
fixed matching block size isn’t suitable for predicting different
kinds of motion

INTRODUCTION

Improved 3DRS Algorithm


BI-DIRECTIONAL BLOCK MATCHING
MULTI-RESOLUTION SEARCHING

EXPERIMENTAL RESULTS

CONCLUSION
Forward Block Matching for Original 3DRS

For each block B(x) at block position x of fn,
the match error criterion is the Sum of Absolute Difference
(SAD) which can be described as follow:



leads to the interpolation
errors propagating into
subsequent output frames
d : motion vector, d  (d x , d y )
for x’ : y mod 2 = n mod 2
BI-DIRECTIONAL BLOCK MATCHING


assumption that the motion at forward and backward
fields is consistent
For each block B(x) at position x of fn, the match error
criterion for bi-directional block matching is as follows:

d : motion vector,
d  (d x , d y )
, the restriction : dy mod
2=0
MULTI-RESOLUTION SEARCHING

Larger block size makes the motion estimation


less sensitive to noise
reduces the aperture problem
not suitable to small-area motion details
Adjust block size according to the video content.
 First, at the lowest resolution, matching block sized 16x16
 By split criterion, decide
whether to go down to a


higher resolution, four 8x8

smallest block size 4x4.
Split Criterion(1/2)


Observation:
A moving object from a long distance (at a low resolution)


It may contain many types of motion.
Pay attention to a part of it (at a high resolution)

You will get a series of similar motions.
Split Criterion(2/2)

formalized as follow:
n x n block
3DRS motion estimation


Function f returns true or false
depending on the similarity
of 9 MVs
get MV
At least 5MVs=MV5?
true
false
split to four
(n/2)x(n/2) block
higher resolution
Candidate Set at Different Resolution(1/2)

Original 3DRS construct a candidate set

spatial candidates, temporal candidates and update candidates
Candidate Set at Different Resolution(2/2)

Improved 3DRS algorithm adopts different compositions
increase horizontal
search range
16x16
8x8
4x4

INTRODUCTION

Improved 3DRS Algorithm


Bi-directional Block Matching
Multi-resolution Searching

EXPERIMENTAL RESULTS

CONCLUSION
Subjective Visual Quality(1/2)

Step1: extract interlaced videos from progressive
sequences by dropping one field in each frame

Step2 : de-interlace the generated videos using



original 3DRS algorithm
improved 3DRS algorithm , respectively.
Video sequence:
1.
2.
Paris
Mobile & calendar
Subjective Visual Quality(2/2)
1.
Video sequence : paris
a) Original 3DRS
2.
b) improved 3DRS
Video sequence : mobile & calender
a) Bi3DRS
b) improved 3DRS
Objective PSNR Performance(1/2)

7 CIF progressive sequences and 2 HD progressive
sequences are selected.

test 3 de-interlacing algorithms:

original 3DRS, Bi3DRS and improved 3DRS algorithm.

each of them is followed by the AR error protection[11]
Objective PSNR Performance(2/2)


the parameter C1 of AR error protection determines the ratio
of temporal and spatial information
If parameter C1 ↑


spatial information ↑ => the impact of error propagation ↓
benefits from accurate motion estimation ↓
CONCLUSION

3DRS algorithm is an efficient motion estimation method
for the video de-interlacing process

Improve 3DRS algorithm to overcome some deficiencies

bi-directional block matching


multi-resolution searching


to eliminate error propagation
to adjust matching block size according to the video content.
Experiments show that better subjective and objective
results than the original 3DRS algorithm on most tested
sequences.