Yen-Lin Lee and Truong Nguyen ECE Dept., UCSD, La Jolla, CA
92093-0407 Method and Architecture Design for Motion Compensated
Frame Interpolation in High-Definition Video Processing
Slide 2
Motion compensated frame interpolation (MCFI) Enhance the
quality of reconstructed video Work on how to accurately estimate
true motions Considers spatial and temporal correlation, searches
with block matching algorithm (BMA) proposed method and
architecture adopt one-pass and low complexity design
Slide 3
The proposed method(1/2) Adopts a motion-compensated approach
to insert one or several interpolated frames between any two
contiguous original frames
Slide 4
The proposed method(2/2) Three major parts: True motion engine
Motion estimation, true motion decision, output true MV to
interpolator Block-based interpolator obtains motion information
performs a motion-compensated procedure with weighted values
Deblocking filter Smooth the blocking artifact where a neighboring
interpolated block with larger sum of absolute difference (SAD)
shows up
Slide 5
Bilateral Search Adopt motion re-estimation with bilateral
search with original uncompressed frames from a decoder directly
seeks two similar blocks from a zero distance of the interpolated
block
Slide 6
Multi-directional Enlarged Matching Algorithm Larger matching
block size could improve the accuracy of searching true motion (
N*N M*M ) larger part of a particular object is considered. Nine
types of enlarged matching directions are defined keep the
completeness of the front edge for the moving object.
Slide 7
One-pass Process with Multi-grids Classification Simplify the
process and reduce the storage size for motion vector field Divides
a searching window into multiple areas Each area has one motion
candidate.
Slide 8
Temporal Object Information One-pass processing would not
provide complete info. acquire some temporal information For
objects with constant speed, especially in panning scenes. When the
camera pans slowly across a scene, most interpolated would move
consistently except real moving objects. Remove those inaccurate
motion vectors Interpolated blocks with a larger pixel variance are
considered, which means that there are more textures.
Slide 9
THE PROPOSED ARCHITECTURE white color : functional elements
gray color : storage elements
Slide 10
True Motion Engine(1/3) it sets nine-grids classification,
searching window from -7 to +7 for a down-sampled frame, 88 block
size, 1616 enlarged block size, and four candidates for each MSEA
sub- block.
Slide 11
True Motion Engine(2/3) Multi-level successive eliminate
algorithm (MSEA) A fast full-search block matching algorithm
(FFSBMA) Following a successive elimination algorithm (SEA) Avoid
local trap problem
Slide 12
True Motion Engine(3/3) The fourth step : Probable sum of
absolute difference (PSAD) MSEA engine decides four highly possible
candidates for each directional area PSAD engine only needs to
check SAD values for these four vectors instead of a full search
with SAD. The fourth step : true motion selection from these nine
motion candidates by pre-defines conditions, and neighboring and
temporal information. The fifth step : The proposed architecture
would perform a refining procedure by probable refining engine
(PREF)
Slide 13
Implementation double the frame rate for : HDTV1080p (19201080)
30fps video at 180MHz HDTV 720p (1280720) 30fps video at 83MHz the
architecture is designed in VHDL and implemented with TSMC 90-nm
technology
Slide 14
PERFORMANCE AND EXPERIMENTAL RESULT(1/5)
Slide 15
PERFORMANCE AND EXPERIMENTAL RESULT(2/5)
Slide 16
PERFORMANCE AND EXPERIMENTAL RESULT(3/5)
Slide 17
PERFORMANCE AND EXPERIMENTAL RESULT(4/5)
Slide 18
PERFORMANCE AND EXPERIMENTAL RESULT(5/5) More result:
http://videoprocessing.ucsd.edu/~yenlinlee/mcfi/
Slide 19
CONCLUSION The method employs a unique true motion engine with
an adaptive overlapped block matching algorithm multi-directional
enlarged matching algorithm one-pass process. The proposed
architecture employs a modified multi-level successive eliminate
algorithm has capabilities to reduce the heavy computation.
Experimental results show that the proposed algorithm provides
better video quality than conventional methods