1 / 15

Quantization

Quantization. Quantization. Signal x(t) is quantized in a finite number of levels Assume that x is in the dynamic range [-1,1[ and we quantize it with b+1 bits -> 2 b+1 levels Quantization introduces an error : e = Q[x] - x. Rounding versus truncation. Truncation: -2 -b < Q[x] – x <= 0

rocco
Télécharger la présentation

Quantization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Quantization

  2. Quantization • Signal x(t) is quantized in a finite number of levels • Assume that x is in the dynamic range [-1,1[ and we quantize it with b+1 bits -> 2b+1 levels • Quantization introduces an error: e = Q[x] - x

  3. Rounding versus truncation Truncation: -2-b < Q[x] – x <= 0 Example: 010111 -> 0101 Rounding: -2-b/2 < Q[x] – x <= 2-b/2 Example: 010111 ->0110

  4. Statistical model • Additive error model: The quantized signal is the sum of the original signal and a quantization noise signal e(t)

  5. Fixed point representation • Two’s complement fractional representation: • Dynamic range: • Advantage of fractional: Multiplier can not overflow(except for (-1).(-1))

  6. Two’s complement fractional • Example b=3

  7. Two’s complement arithmetic (1) • Addition: • Addition can lead to an overflow. Need to scale the input. • If no overflow x+y can be represented exactly with b+1 bits. Example • In a filter check the scaling (max/min signal values) at the output of each adder. • An adder does not inject quantization noise

  8. Two’s complement arithmetic's (2) • Multiplication: • If coefficient is not –1, multiplier can not overflow • x,y represented with b+1 bits -> x.y is exactly represented with 2b+1 bits Example: • If the multiplier output is forced into a register of lenght < 2b+1 it introduces a quantization noise whose max amplitude is 1 lsb (truncation) or +- ½ lsb (rounding).

  9. Multiplier model • If b3 <b1+b2-1 the truncation or rounding of the multiplier output is modeled as an additive noise e(n)

  10. First-order IIR example • Feedback path -> register size at multiplier output can not be increased at each iteration! Problem with IIR filters • Noise e2(n) is amplified by the feedback loop (a is close to 1). We will choose b2 >b1 • Filter design: compute the PSD at the filter output due to each quantization noise. Statistics on e(n) needed

  11. First order statistics (1) • e(n) is a random process = signal that can, for each time index n, be described as a random variable e • Probability Density Function pe(u) pe(u).De = probability that the random variable e be in the interval [u,u+Du] • Uniform PDF: pe(u) is constant over a range W and zero elsewhere. Recall that the total area underneath pe(u) must integrate to 1 Truncation Rounding

  12. First order statistics (2) • The hypothesis that the quantization error is uniform does not hold if the input signal covers less than 1 lsb! It is assumed valid if the signals in all registers of the filter cover a significant portion of the register dynamic range. • Mean: DC component of the noise • Variance: total AC power of the noise

  13. Second order statistics • Power Spectral Density of a random process e(n): See(ejω) d ω = power in the band d ω centered at ω In z domain, power spectral density PSD when there is only one noise input is defined as: PSD, when there are K noise inputs, is

  14. Quantization noise • One noise source e(n) is fed into a linear filter and the output is f(n) • Mean: • Power Spectrum:

  15. Quantization noise • Variance: • Special case if e(n) is white We will assume that the quantization noise e(n) is white

More Related