Transform Coding. Transform Coding Principle

Similar documents
Transform Coding. Transform Coding Principle

Transform coding - topics. Principle of block-wise transform coding

VQ widely used in coding speech, image, and video

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Lossy Compression. Compromise accuracy of reconstruction for increased compression.

Pulse Coded Modulation

Flexible Quantization

Composite Hypotheses testing

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

Error Probability for M Signals

Tutorial 2. COMP4134 Biometrics Authentication. February 9, Jun Xu, Teaching Asistant

Lecture 3: Shannon s Theorem

Chapter 7 Channel Capacity and Coding

Fourier Transform. Additive noise. Fourier Tansform. I = S + N. Noise doesn t depend on signal. We ll consider:

Chapter 8 SCALAR QUANTIZATION

A NEW DISCRETE WAVELET TRANSFORM

STATISTICAL DISTRIBUTIONS OF DISCRETE WALSH HADAMARD TRANSFORM COEFFICIENTS OF NATURAL IMAGES

Chapter 7 Channel Capacity and Coding

Lec 12 Rate-Distortion Optimization (RDO) in Video Coding-II

Communication with AWGN Interference

Digital Signal Processing

EGR 544 Communication Theory

Report on Image warping

COMPUTATIONALLY EFFICIENT WAVELET AFFINE INVARIANT FUNCTIONS FOR SHAPE RECOGNITION. Erdem Bala, Dept. of Electrical and Computer Engineering,

Entropy Coding. A complete entropy codec, which is an encoder/decoder. pair, consists of the process of encoding or

Digital Modems. Lecture 2

Lec 07 Transforms and Quantization II

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Lec 13 Review of Video Coding

GEMINI GEneric Multimedia INdexIng

Some Comments on Accelerating Convergence of Iterative Sequences Using Direct Inversion of the Iterative Subspace (DIIS)

Lec 13 Review of Video Coding

What would be a reasonable choice of the quantization step Δ?

Lossless Compression Performance of a Simple Counter- Based Entropy Coder

Asymptotic Quantization: A Method for Determining Zador s Constant

A linear imaging system with white additive Gaussian noise on the observed data is modeled as follows:

Natural Images, Gaussian Mixtures and Dead Leaves Supplementary Material

CSE4210 Architecture and Hardware for DSP

Why Bayesian? 3. Bayes and Normal Models. State of nature: class. Decision rule. Rev. Thomas Bayes ( ) Bayes Theorem (yes, the famous one)

DFT with Planewaves pseudopotential accuracy (LDA, PBE) Fast time to solution 1 step in minutes (not hours!!!) to be useful for MD

The Gaussian classifier. Nuno Vasconcelos ECE Department, UCSD

Multigradient for Neural Networks for Equalizers 1

A Self-embedding Robust Digital Watermarking Algorithm with Blind Detection

Lecture Notes on Linear Regression

Statistical pattern recognition

Lecture 3: Dual problems and Kernels

Novel Pre-Compression Rate-Distortion Optimization Algorithm for JPEG 2000

Conversion Scheme for DCT-domain Transcoding of MPEG-2 to H.264/AVC

A Fast Fractal Image Compression Algorithm Using Predefined Values for Contrast Scaling

Memory ecient adaptation of vector quantizers to time-varying channels

Conversion Scheme for DCT-Domain Transcoding of MPEG-2 to H.264/AVC

Lecture 14 (03/27/18). Channels. Decoding. Preview of the Capacity Theorem.

Generalized Linear Methods

Linear Approximation with Regularization and Moving Least Squares

Gaussian Conditional Random Field Network for Semantic Segmentation - Supplementary Material

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

Switched Quasi-Logarithmic Quantizer with Golomb Rice Coding

ECE 472/572 - Digital Image Processing. Roadmap. Questions. Lecture 6 Geometric and Radiometric Transformation 09/27/11

COMPARING NOISE REMOVAL IN THE WAVELET AND FOURIER DOMAINS

Lecture 12: Classification

Invariant deformation parameters from GPS permanent networks using stochastic interpolation

Fingerprint Enhancement Based on Discrete Cosine Transform

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Application of Nonbinary LDPC Codes for Communication over Fading Channels Using Higher Order Modulations

Chapter 11: Simple Linear Regression and Correlation

Dynamic Programming. Preview. Dynamic Programming. Dynamic Programming. Dynamic Programming (Example: Fibonacci Sequence)

Math Review. CptS 223 Advanced Data Structures. Larry Holder School of Electrical Engineering and Computer Science Washington State University

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

Time-Varying Systems and Computations Lecture 6

Multi-dimensional Central Limit Argument

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Convergence of random processes

Antenna Combining for the MIMO Downlink Channel

Consider the following passband digital communication system model. c t. modulator. t r a n s m i t t e r. signal decoder.

Johann Wolfgang Goethe-Universität Frankfurt am Main

3.1 ML and Empirical Distribution

Exercises. 18 Algorithms

Mixed Noise Suppression in Color Images by Signal-Dependent LMS L-Filters

Assuming that the transmission delay is negligible, we have

OPTIMAL COMBINATION OF FOURTH ORDER STATISTICS FOR NON-CIRCULAR SOURCE SEPARATION. Christophe De Luigi and Eric Moreau

The Prncpal Component Transform The Prncpal Component Transform s also called Karhunen-Loeve Transform (KLT, Hotellng Transform, oregenvector Transfor

Signal space Review on vector space Linear independence Metric space and norm Inner product

Estimating the Fundamental Matrix by Transforming Image Points in Projective Space 1

On MIMO Channel Capacity with Output Quantization Constraints

Power Allocation for Distributed BLUE Estimation with Full and Limited Feedback of CSI

Quantum and Classical Information Theory with Disentropy

P R. Lecture 4. Theory and Applications of Pattern Recognition. Dept. of Electrical and Computer Engineering /

Lecture 4: Universal Hash Functions/Streaming Cont d

S Advanced Digital Communication (4 cr) Targets today

Limited Dependent Variables

ECE 534: Elements of Information Theory. Solutions to Midterm Exam (Spring 2006)

Integrating Neural Networks and PCA for Fast Covert Surveillance

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Statistical analysis using matlab. HY 439 Presented by: George Fortetsanakis

Comparison of Wiener Filter solution by SVD with decompositions QR and QLP

LECTURE 9 CANONICAL CORRELATION ANALYSIS

Multi-dimensional Central Limit Theorem

Transcription:

Transform Codng Prncple of block-wse transform codng Propertes of orthonormal transforms Dscrete cosne transform (DCT) Bt allocaton for transform coeffcents Entropy codng of transform coeffcents Typcal codng artfacts Fast mplementaton of the DCT Thomas Wegand: Dgtal Image Communcaton Transform Codng - Transform Codng Prncple Structure u Τ U Quantzer V Τ v Transform coder (Τα) / decoder (βτ ) structure u U V Τ α β Τ v Insert entropy codng (γ) and transmsson channel u Τ U b b V α γ channel γ β Τ v Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Transform Codng and Quantzaton u U V v u Τ U V Q U V Τ u v v Transformaton of vector u=(u,u,,u )ntou=(u,u,,u ) Quantzer Q maybe appled to coeffcents U separately (scalar quantzaton: low complexty) jontly (vector quantzaton, may requre hgh complexty, explotng of redundancy between U ) Inverse transformaton of vector V=(V,V,,V )ntov=(v,v,,v ) Why should we use a transform? Thomas Wegand: Dgtal Image Communcaton Transform Codng - 3 Geometrcal Interpretaton A lnear transform can decorrelate random varables An orthonormal transform s a rotaton of the sgnal vector around the orgn Parseval s Theorem holds for orthonormal transforms DC bass functon U u U u u u = = T = u AC bass functon U U 3 Tu U = = = = Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Propertes of Orthonormal Transforms Forward transform transform coeffcents U=Tu Transform matrx of sze x nput sgnal block of sze Orthonormal transform property: nverse transform u=t U=TT U Lnearty: u s represented as lnear combnaton of bass functons Thomas Wegand: Dgtal Image Communcaton Transform Codng - 5 Transform Codng of Images Explot horzontal and vertcal dependences by processng blocks orgnal mage reconstructed mage Orgnal mage block Reconstructed mage block Transform T Transform coeffcents Quantzaton Entropy Codng Transmsson Inverse transform T - Quantzed transform coeffcents Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Separable Orthonormal Transforms, I Problem: sze of vectors * (typcal value of : 8) An orthonormal transform s separable, f the transform of a sgnal block of sze *-can be expressed by U=TuT * transform coeffcents Orthonormal transform matrx of sze * * block of nput sgnal The nverse transform s u= TT UT Great practcal mportance: transform requres matrx multplcatons of sze * nstead one multplcaton of a vector of sze* wthamatrxofsze * Reducton of the complexty from O( )too( 3 ) Thomas Wegand: Dgtal Image Communcaton Transform Codng - 7 Separable Orthonormal Transforms, II Separable -D transform s realzed by two -D transforms along rows and columns of the sgnal block u column-wse Tu row-wse TuTT -transform -transform * block of pxels * block of transform coeffcents Thomas Wegand: Dgtal Image Communcaton Transform Codng - 8

Crtera for the Selecton of a Partcular Transform Decorrelaton, energy concentraton KLT, DCT, Transform should provde energy compacton Vsually pleasant bass functons pseudo-random-nose, m-sequences, lapped transforms, Quantzaton errors make bass functons vsble Low complexty of computaton Separablty n -D Smple quantzaton of transform coeffcents Thomas Wegand: Dgtal Image Communcaton Transform Codng - 9 Karhunen Loève Transform (KLT) Decorrelate elements of vector u R u =E{uu T }, U=Tu, R U =E{UU T }=TE{uu T }T T = TR u T T =dag{α } Bass functons are egenvectors of the covarance matrx of the nput sgnal. KLT acheves optmum energy concentraton. Dsadvantages: - KLT dependent on sgnal statstcs - KLT not separable for mage blocks - Transform matrx cannot be factored nto sparse matrces. Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Comparson of Varous Transforms, I Comparson of D bass functons for block sze =8 () Karhunen Loève transform (98/9) () Haar transform (9) (3) Walsh-Hadamard transform (93) () Slant transform (Enomoto, Shbata, 97) () () (3) () (5) (5) Dscrete CosneTransform (DCT) (Ahmet, atarajan, Rao, 97) Thomas Wegand: Dgtal Image Communcaton Transform Codng - Comparson of Varous Transforms, II Energy concentraton measured for typcal natural mages, block sze x3 (Lohscheller) KLT s optmum DCT performs only slghtly worse than KLT Thomas Wegand: Dgtal Image Communcaton Transform Codng -

DCT - Type II-DCT of blocksze M x M - D bass functons of the DCT: s defned by transform matrx A contanng elements π (k + ) ak = α cos M k, =...( M ) wth α = M α = M Thomas Wegand: Dgtal Image Communcaton Transform Codng - 3 Dscrete Cosne Transform and Dscrete Fourer Transform Transform codng of mages usng the Dscrete Fourer Transform (DFT): edge For statonary mage statstcs, the energy folded concentraton propertes of the DFT converge aganst those of the KLT for large block szes. Problem of blockwse DFT codng: blockng effects DFT of larger symmetrc block -> DCT due to crcular topology of the DFT and Gbbs phenomena. pxel folded Remedy: reflect mage at block boundares Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Hstograms of DCT Coeffcents: Image: Lena, 5x5 pxel DCT: 8x8 pxels DCT coeffcents are approxmately dstrbuted lke Laplacan pdf Thomas Wegand: Dgtal Image Communcaton Transform Codng - 5 Dstrbuton of the DCT Coeffcents Central Lmt Theorem requests DCT coeffcents to be Gaussan dstrbuted Model varance of Gaussan DCT coeffcents dstrbuton as random varable (Lam & Goodmann ) u pu ( σ ) = exp πσ σ Usng condtonal probablty pu ( ) = pu ( σ ) p( σ ) dσ Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Dstrbuton of the Varance Exponental Dstrbuton { } p( σ ) = µ exp µσ Thomas Wegand: Dgtal Image Communcaton Transform Codng - 7 Dstrbuton of the DCT Coeffcents pu ( ) = pu ( σ ) p( σ ) dσ u = exp µ exp{ µσ } dσ πσ σ u = µ exp µσ dσ π σ π µ u = µ exp π µ µ = exp { µ u } Laplacan Dstrbuton Thomas Wegand: Dgtal Image Communcaton Transform Codng - 8

Bt Allocaton for Transform Coeffcents I Problem: dvde bt-rate R among transform coeffcents such that resultng dstorton D s mnmzed. DR ( ) = D ( R ), s.t. R R = = Average Average dstorton rate Dstorton contrbuted by coeffcent Rate for coeffcent Approach: mnmze Lagrangan cost functon d dd ( )! R D( R) + λ R = + λ= dr dr = = Soluton: Pareto condton dd( R) = λ dr Move bts from coeffcent wth small dstorton reducton per bt to coeffcent wth larger dstorton reducton per bt Thomas Wegand: Dgtal Image Communcaton Transform Codng - 9 Bt Allocaton for Transform Coeffcents II Assumpton: hgh rate approxmatons are vald dd ( R ) D R aσ a σ = λ R R ( ), ln dr aln R σ logσ + log R logσ logσ R log R λ + = σ + ( σ ) = = aln ln logσ + log = log a R R + log = = λ = λ log σ σ Operatonal Dstorton Rate functon for transform codng: DR ( ) = D( R) = = a σ σ σ σ log a R R a R σ = = = = Thomas Wegand: Dgtal Image Communcaton Transform Codng - Geometrc mean: σ = ( σ )

Entropy Codng of Transform Coeffcents Prevous dervaton assumes: AC coeffcents are very lkely to be zero Orderng of the transform coeffcents by zg-zag-scan Desgn Huffman code for event: # of zeros and coeffcent value Arthmetc code maybe smpler Probablty that coeffcent s not zero when quantzng wth: V =*round(u /).5 R 3 σ max[(log ),] = D 5 7 8 bt DC coeffcent 8 7 5 3 Thomas Wegand: Dgtal Image Communcaton Transform Codng - Entropy Codng of Transform Coeffcents II Encoder 95 88 93 9 57 9 9 93 88 87 95 93 3 93 8 9 8 95 8 5 99 93 7 7 79 79 5 8 98 83 9 95 9 7 59 85 8 75 3 5 7 73 85 5 7 5 7 8 8 7 73 98 3 5 8 9 9 59 3 Orgnal 8x8 block DCT 8 9 33-5 - 33-38 -5-7 -3-9 3 - - - 9 8 7 7-3 -5-3 - 3-8 -3-3 -5-9 3 3 3-7 - - -9 8-7 -5-8 -7 α 85 3-3 - - - - - - - zg-zag-scan (85 3 - -3 - - - - - - EOB) run-level codng Mean of block: 85 (,3) (,) (,) (,) (,) (,-) (,) (,) (,) (,-3) (,) (,-) (,) (,-) (,-) (,-) (,) (9,-) (,-) (EOB) entropy codng & transmsson Thomas Wegand: Dgtal Image Communcaton Transform Codng -

3 - - -3 3 - - -3 3 - - -3 3 - - -3 3 - - -3 3 - - -3 Entropy Codng of Transform Coeffcents III Decoder transmsson Mean of block: 85 (,3) (,) (,) (,) (,) (,- ) (,) (,) (,) (,-3) (,) (,-) (,) (,-) (,-) (,-) (,) (9,-) (,-) (EOB) Inverse zg-zagscan run-leveldecodng (85 3 - -3 - - - - - - EOB) 85 3-3 - - - - - - - Scalng and nverse DCT: β 9 93 87 9 79 7 9 89 98 88 8 98 9 9 8 85 89 9 97 7 59 8 89 7 8 8 77 5 53 87 89 99 78 5 3 85 79 7 93 7 5 79 97 7 9 98 95 93 9 5 8 79 9 9 9 85 9 57 Reconstructed 8x8 block Thomas Wegand: Dgtal Image Communcaton Transform Codng - 3 Detal n a Block vs. DCT Coeffcents Transmtted mage block DCT coeffcents quantzed DCT block of block coeffcents reconstructed of block from quantzed coeffcents Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Typcal DCT Codng Artfacts DCT codng wth ncreasngly coarse quantzaton, block sze 8x8 quantzer step sze quantzer step sze quantzer step sze for AC coeffcents: 5 for AC coeffcents: for AC coeffcents: Thomas Wegand: Dgtal Image Communcaton Transform Codng - 5 Influence of DCT Block Sze Effcency as a functon of blocksze x, measured for 8 bt Quantzaton n the orgnal doman and equvalent quantzaton n the transform doman. = Memoryless entropy of orgnal sgnal mean entropy of transform coeffcents Block sze 8x8 s a good compromse. Thomas Wegand: Dgtal Image Communcaton Transform Codng -

Fast DCT Algorthm I DCT matrx factored nto sparse matrces (Ara, Agu, and akajma; 988): y = M x = S P M M M 3 M M 5 M x S = s s s s3 s s5 s s7 P = M = - - M = - - M = 3 C -C C -C -C C M = - M = 5 - - - - M = - - - - Thomas Wegand: Dgtal Image Communcaton Transform Codng - 7 Fast DCT Algorthm II Sgnal flow graph for fast (scaled) 8-DCT accordng to Ara, Agu, akajma: x s y scalng x s y x x 3 x x 5 a a a 3 s s s 5 s y y y 5 y only 5 + 8 Multplcatons (drect matrx multplcaton: multplcatons) x a x 7 a 5 s 7 s 3 y 7 y 3 Multplcaton: u m m u Addton: u v u v u+v u-v a= C a = C C a3 = C a = C+ C a5 = C s = sk = ; k =... 7 CK CK = cos(ks ) Thomas Wegand: Dgtal Image Communcaton Transform Codng - 8

Transform Codng: Summary Orthonormal transform: rotaton of coordnate system n sgnal space Purpose of transform: decorrelaton, energy concentraton KLT s optmum, but sgnal dependent and, hence, wthout a fast algorthm DCT shows reduced blockng artfacts compared to DFT Bt allocaton proportonal to logarthm of varance Threshold codng + zg-zag-scan + 8x8 block sze s wdely used today (e.g. JPEG, MPEG, ITU-T H.3) Fast algorthm for scaled 8-DCT: 5 multplcatons, 9 addtons Thomas Wegand: Dgtal Image Communcaton Transform Codng - 9