Volodymyr Fedak Artifacts suppression in images and video.

Volodymyr Fedak

Artifacts suppression in images and video

Introduction

What is the problem?

Why is it important?

What did I do? What are the results?

So what next?

What is the problem?

blocking

ringing

blurring

flickering

What is the problem?

F - 2F - 1

F

F + 1F + 2

Intra-frame processing… Inter-frame processing…

Why is it important ?

De-coder Artifact detection

Reducing artifacts

Transform to original format

Enhanced information

postprocessingCoder parameters

Compressed information

Postprocessing techniques:•motion-compensated algorithms

iterative approaches based on the

theory of projections onto convex set •spatial-temporal algorithms

algorithms that transform signal

to frequency domain

What did I do ?

Analyse modern postprocessing techniques

Implement most encouraging methods

Compare results of mentioned algorithms

Propose approaches for optimization

Wavelet-based de-blocking and de-ringing algorithm proposed by Alan and Liew

Steps:•Detection of Block Discontinuities•Threshold Maps Generation at Different Wavelet Scales•low frequency filtering

Non-Local Means

NLM is an improvement of Bilateral filtering

dyxIyIsxycyIxI ))(),((),()()(

C(y, x) - geometric relationship

S(I(y), I(x)) - luminance ratio

I(y) – pixel luminance

Non-Local Means

NLM could be presented:in general way:

)(),())(( jvjiwivNLIj

v(i) – noisy imageW(i, j) - weighted average of pixels in the image v(j) – pixel luminance

in terms of implementation:

2

2)()(

)(

)()(

1 h

yNxN

xQyh exz

xCxNL

)(

)()(

2

2

2

)(xQy

h

yNxN

exC

N(x) - window surrounding pixel x;Q(x) is a search window around pixel x;

Non-Local Means Parameters

•h - determines the amount of averaging (h increases amount of blocking artifacts decrease).•N (x) – the match window/patch – when N(x) increases, blocking artifacts of the processed sequence decreases very slowly •Q(x) – the search window/patch – when Q(x) increases, artifacts of the processed sequence decreases very slowly for an increasing value of the search window size, and we have a large amount of computation time.

Possible ways for optimization:

•Extended NLM to the temporal domain . Use together with motion-compensation algorithm but apply some quality coefficient to the motion vector. •Add smart patch/search window size choosing algorithm.•Use Hierarchical block matching algorithm to find similar windows for speeding-up NLM

Any questions ?