Scalable Wavelet Video Coding Using Aliasing-Reduced Hierarchical Motion Compensation

Xuguang Yang, Member, IEEE, and Kannan Ramchandran, Member, IEEE

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 9, NO. 5, MAY 2000

Outline Introduction Basic derivation Basic system structure Backward/forward hybrid motion compensation Computational complexity Coding results Conclusion and future research

Introduction

T QEntropycoding

Image block

TransformCoefficients

Zigzag Scan(2D->1D)

Bitstream

• Encoder For Video Sequence

Q-1T-1

ReconstructedTransformCoefficients

ReconstructedImage block

MC

-

Aliasing-Reduced motion estimationBackward/forward motion estimation

DWTDWT

DWT-1

Features of wavelet base video coding Support scalability Free from blocky artifacts

Basic Derivation Aliasing come from downsampling.

tπ 2π-2π π

tπ 2π-2π π

tπ 2π-2π π

cos wt fs=2π fs=1/2π

fs=3/2π

fs=3π --- Aliasing

Basic Derivation

Aliasing Problem

Aliasing Reduction Using an Interpolation Filter

Signal preservation

Aliasing reduction

Aliasing Reduction Using an Interpolation Filter

Input power spectral density

Expectation

Optimal Solution

Time domain

Optimal Solution

Basic System Structure

Three Level wavelet transform

Use frame difference coding

Basic System Structure

12

3

4 Q & E

4x4

OBMC

5 Repeat to nextstage

Basic Operations

Backward/forward hybrid motion compensation Reason

Experiments have revealed a degraded performance at low bit rates and very complicated motion.

Accuracy is dependent on the reconstruction quality of coarser frames.

Backward/forward hybrid motion compensation Zerotrees of Mode Selections Mode Optimization Dynamic Programming Algorithm Choose of λm

Zerotrees of Mode Selections

forward

Mode Optimization Initializing all the tree nodes to backward

mode. Mode selection is performed as comparing

the R-D Lagrangian

Bottom-up dynamic programming strategy

Note that the distortion here is the motion compensated error energy, not the final coded distortion.

Dynamic Programming AlgorithmBackward cost

Forward cost

Df < DbForwardBackward

Aggregated Lagrangian gain1

2

Dynamic Programming AlgorithmBackward cost

Forward cost

Df < Db

ForwardBackward

3

A Toy Example

Choice of λm

“Lagrangian compression ratio”

Choice of λm

is almost solely a function of Given a certain

Find Cl(λ) by training, and send it as side information

Dynamic Programming Algorithm

1

2

3

Computational Complexity Great computational savings can be achieved by

taking advantage of the striking similarities between motion vectors in successive resolution levels, and between the backward and forward motion vectors.

The increment is proportional to the square of the ratio between forward search range and backward search range, which is typically 20%–30%.

Computational Complexity The quadtree optimization algorithm.

The total computation for the optimization is therefore of O(N) complexity ( N is the total number of pixels), which is negligible compared to the O(N2) complexity of motion estimation.

While the proposed coder saves complexity at the encoder, it requires an increase in decoder complexity.

Coding Results

Direct estimation

Interpolated estimation using the synthesis lowpass filter G0(w)

Interpolated estimation using the L(w)

The backward motion compensated error energy on 100 frames of the football sequence at three resolution levels

Final coded PSNR for luminance versus frame number at 15 frames/sUse L(w)

Use G0(w)H.263 with full option

MaD48kb/s, 15fs

Missa24kb/s, 15fs

0.5-1.5dB over H.263Average 0.87dB

Final coded PSNR for luminance versus frame number at 15 frames/s

Final coded PSNR for luminance versus frame number at 30 frames/s

MPEG-2

Propose method

Final coded PSNR for luminance versus frame number at 30 frames/s

Scalable decoding

0.5Mb/s

Comparison of final coded subjective quality

H.263 at 48 Kb/s Proposed Coder

“mosquito” noise ?

Comparison of final coded subjective quality

H.263 at 24 Kb/s Proposed Coder

“mosquito” noise ?

Comparison of final coded subjective quality

MPEG-2 at 2Mb/s Proposed Coder

Conclusions and future research Proposed coder alleviates the aliasing probl

em in motion estimation. Backward/forward hybrid motion compens

ation attack the instability problem caused by quantization noise. (2dB)

Spatially scalable. Ringing effects as a result of wavelet transf

orm coding.