Fundamentals of Vector Quantization

Embed Size (px)

Citation preview

  • 8/18/2019 Fundamentals of Vector Quantization

    1/87

  • 8/18/2019 Fundamentals of Vector Quantization

    2/87

  • 8/18/2019 Fundamentals of Vector Quantization

    3/87

  • 8/18/2019 Fundamentals of Vector Quantization

    4/87

  • 8/18/2019 Fundamentals of Vector Quantization

    5/87

  • 8/18/2019 Fundamentals of Vector Quantization

    6/87

  • 8/18/2019 Fundamentals of Vector Quantization

    7/87

  • 8/18/2019 Fundamentals of Vector Quantization

    8/87

  • 8/18/2019 Fundamentals of Vector Quantization

    9/87

  • 8/18/2019 Fundamentals of Vector Quantization

    10/87

  • 8/18/2019 Fundamentals of Vector Quantization

    11/87

  • 8/18/2019 Fundamentals of Vector Quantization

    12/87

  • 8/18/2019 Fundamentals of Vector Quantization

    13/87

  • 8/18/2019 Fundamentals of Vector Quantization

    14/87

  • 8/18/2019 Fundamentals of Vector Quantization

    15/87

  • 8/18/2019 Fundamentals of Vector Quantization

    16/87

  • 8/18/2019 Fundamentals of Vector Quantization

    17/87

  • 8/18/2019 Fundamentals of Vector Quantization

    18/87

  • 8/18/2019 Fundamentals of Vector Quantization

    19/87

  • 8/18/2019 Fundamentals of Vector Quantization

    20/87

  • 8/18/2019 Fundamentals of Vector Quantization

    21/87

  • 8/18/2019 Fundamentals of Vector Quantization

    22/87

  • 8/18/2019 Fundamentals of Vector Quantization

    23/87

  • 8/18/2019 Fundamentals of Vector Quantization

    24/87

  • 8/18/2019 Fundamentals of Vector Quantization

    25/87

  • 8/18/2019 Fundamentals of Vector Quantization

    26/87

  • 8/18/2019 Fundamentals of Vector Quantization

    27/87

  • 8/18/2019 Fundamentals of Vector Quantization

    28/87

  • 8/18/2019 Fundamentals of Vector Quantization

    29/87

  • 8/18/2019 Fundamentals of Vector Quantization

    30/87

  • 8/18/2019 Fundamentals of Vector Quantization

    31/87

  • 8/18/2019 Fundamentals of Vector Quantization

    32/87

  • 8/18/2019 Fundamentals of Vector Quantization

    33/87

  • 8/18/2019 Fundamentals of Vector Quantization

    34/87

  • 8/18/2019 Fundamentals of Vector Quantization

    35/87

  • 8/18/2019 Fundamentals of Vector Quantization

    36/87

  • 8/18/2019 Fundamentals of Vector Quantization

    37/87

  • 8/18/2019 Fundamentals of Vector Quantization

    38/87

  • 8/18/2019 Fundamentals of Vector Quantization

    39/87

  • 8/18/2019 Fundamentals of Vector Quantization

    40/87

  • 8/18/2019 Fundamentals of Vector Quantization

    41/87

  • 8/18/2019 Fundamentals of Vector Quantization

    42/87

  • 8/18/2019 Fundamentals of Vector Quantization

    43/87

  • 8/18/2019 Fundamentals of Vector Quantization

    44/87

  • 8/18/2019 Fundamentals of Vector Quantization

    45/87

  • 8/18/2019 Fundamentals of Vector Quantization

    46/87

  • 8/18/2019 Fundamentals of Vector Quantization

    47/87

  • 8/18/2019 Fundamentals of Vector Quantization

    48/87

  • 8/18/2019 Fundamentals of Vector Quantization

    49/87

  • 8/18/2019 Fundamentals of Vector Quantization

    50/87

  • 8/18/2019 Fundamentals of Vector Quantization

    51/87

  • 8/18/2019 Fundamentals of Vector Quantization

    52/87

  • 8/18/2019 Fundamentals of Vector Quantization

    53/87

  • 8/18/2019 Fundamentals of Vector Quantization

    54/87

  • 8/18/2019 Fundamentals of Vector Quantization

    55/87

  • 8/18/2019 Fundamentals of Vector Quantization

    56/87

  • 8/18/2019 Fundamentals of Vector Quantization

    57/87

  • 8/18/2019 Fundamentals of Vector Quantization

    58/87

  • 8/18/2019 Fundamentals of Vector Quantization

    59/87

  • 8/18/2019 Fundamentals of Vector Quantization

    60/87

  • 8/18/2019 Fundamentals of Vector Quantization

    61/87

  • 8/18/2019 Fundamentals of Vector Quantization

    62/87

  • 8/18/2019 Fundamentals of Vector Quantization

    63/87

  • 8/18/2019 Fundamentals of Vector Quantization

    64/87

  • 8/18/2019 Fundamentals of Vector Quantization

    65/87

  • 8/18/2019 Fundamentals of Vector Quantization

    66/87

  • 8/18/2019 Fundamentals of Vector Quantization

    67/87

  • 8/18/2019 Fundamentals of Vector Quantization

    68/87

  • 8/18/2019 Fundamentals of Vector Quantization

    69/87

  • 8/18/2019 Fundamentals of Vector Quantization

    70/87

  • 8/18/2019 Fundamentals of Vector Quantization

    71/87

  • 8/18/2019 Fundamentals of Vector Quantization

    72/87

  • 8/18/2019 Fundamentals of Vector Quantization

    73/87

  • 8/18/2019 Fundamentals of Vector Quantization

    74/87

  • 8/18/2019 Fundamentals of Vector Quantization

    75/87

  • 8/18/2019 Fundamentals of Vector Quantization

    76/87

  • 8/18/2019 Fundamentals of Vector Quantization

    77/87

  • 8/18/2019 Fundamentals of Vector Quantization

    78/87

  • 8/18/2019 Fundamentals of Vector Quantization

    79/87

    If N very large, then lattice codes chosen so thatVoronoi cell has minimum normalized moment of

    inertia are nearly optimal in the ECVQ problem.Tradeoff: Lattice codes mean simple NN selectionand nearly optimal variable rate performance,but need to entropy code to attain performance.ECVQ design can do better, and sometimesmuch better if rate not asymptotically big.

    Note:

    •For the xed rate case, cells should haveroughly equal partial distortion.•For the variable rate case, cells should haveroughly equal volume.

    In neither case should you try to make cells haveequal probability (maximum entropy).

    79

  • 8/18/2019 Fundamentals of Vector Quantization

    80/87

    Summary

    D(Q) ≈E [ m(X )λ(X )2/k ]N −

    2/k

    D̂k,f (R) ≈C k||f X ||k/ (k+2) 2−2R

    D̂k,v(R)

    ≈C k2

    2h(X )k 2−2R

    Can use high rate quantization theory toquantify the gains of vector quantization overscalar quantization as a function of dimension k,and to separate out the gains as those duememory, to space lling (moment of inertia), andto shape (of density function).

    80

  • 8/18/2019 Fundamentals of Vector Quantization

    81/87

    Gish & Pierce

    Asymptotic theory implies that for iid sources,high rates (low distortion)

    D̂1,v(R)D̂k,v(R) ≤

    C 1C ∞≈

    1.533dB

    Or, equivalently, for low distortionR̂1,v(D) − R̂k,v(D) ≤0.254 bits

    famous “quarter bit” result.

    Suggests at high rates there may be little to begained by using vector quantization (but stillneed to use vectors to do entropy coding!)

    Ziv (1985) showed that for all distortions

    R̂1,v

    (D)

    − R̂

    k,v(D)

    ≤0.754 bits

    using a dithering argument: X̂ = Q(X + Z ) −Z with Z uniform and independent of X .(subtractive dither)

    81

  • 8/18/2019 Fundamentals of Vector Quantization

    82/87

    Comment: If x R and let k → ∞, thenD̂

    k,v(R) →

    k → ∞D(R)

    D̂k,f (R) →k → ∞D(R)

    the Shannon DRF.

    Bennett theory recently extended toinput-weighted squared error measures.

    82

  • 8/18/2019 Fundamentals of Vector Quantization

    83/87

    Recent Extensions

    Can generalize to input-dependent squared errorand to non-difference distortion measures thatbehave locally in this way (have a well-behavedTaylor series expansion)

    d(X, ˆX ) = (X −

    ˆX )∗BX (X −

    ˆX ),Fixed rate:

    D ≥DL(Qopt) = kk + 2

    C −2k

    k N −2k×

    [f (x)(det(B (x)))1k]

    11+2/k dx

    }k+2

    k .Optimal point density

    λ(x)∝( p(x)(det(B (x)))1k)

    kk+2 .

    Variable rate:

    D ≥ DL(Qopt)=

    kC −2k

    kk + 2 ·e−

    2k(H Q−h( p)−12 log(det(B(x)))f (x)dx)

    83

  • 8/18/2019 Fundamentals of Vector Quantization

    84/87

    λopt

    (x) = (det(B (x)))

    12

    x∈G(det(B (x)))12dx

    . (4)

    84

  • 8/18/2019 Fundamentals of Vector Quantization

    85/87

    Final Comments

    •The Shannon compression model:codes, rate, and distortion Mathematical model of compression system.Successfully applied to asymptoticperformance bounds and design algorithms.

    •Optimalility properties and quantizer design Lloyd optimal codes. For high rate, latticecodes + lossless codes nearly optimal (variablerate)Rate-distortion ideas can be used to improvestandards-compliant and wavelet codes. (e.g.,Ramchandran, Vetterli, Orchard)Multiple distortion measures (include othersignal processing)MPEG 4 style image structure?

    85

  • 8/18/2019 Fundamentals of Vector Quantization

    86/87

    Compression on networks, joint source andchannel coding

    •Structured vector quantizers Transform/pyramid/subband/wavelet arecurrently best (rate, distortion, complexity)for image compression, especially embeddedzerotrees

    Many variations and alternative structurescome and go, some stay.

    •Optimal achievable performance:quantization and source coding theory Very general Shannon theorems exist. Recenturry of work on universal and adaptive codes.Multiple distortion measures, non-difference(perceptual) distortion measures.Current work on convergence and rate of convergence (Vapnic-Cernovenkis theory), usein classication and regression (Devroye,Gyor, Gabor, Nobel, Olshen)

    86

  • 8/18/2019 Fundamentals of Vector Quantization

    87/87

    Open problems: Gersho’s conjecture, correctcombination of wavelet theory + Bennett

    (Goldberg did for traditional bit allocationapproach, not yet done for embedded zerotreecodes and related), other improvedcombinations of wavelet and compressiontheory and design (Mallat, Orchard).

    Another issue: Perceived quality of compressedaudio and images.Different applications have different fundamentalquality requirements:Entertainment, browsing, screening, diagnostic,legal, scienticQuantitative as predictors for subjective anddiagnostic.