Sie sind auf Seite 1von 29

Noise INF 5440 - CMOS Image Sensors

NOISE

AO 10V 5.1
Noise INF 5440 - CMOS Image Sensors

Noise Categories

Ref: Nakamura

AO 10V 5.2
Noise INF 5440 - CMOS Image Sensors

Fixed Pattern Noise Output signal


FPN (fixed pattern noise) in dark images
Max level (white)
= DSNU (Dark Signal Non-Uniformity)
• Variations in offset and gain
Constant deviations between read out channels offset and gain.
• Pixel wise deviations in dark currents Dark level
(dark signal increases with the integration time)
Crystal defects Pixels in a row or column
Surface effects related to the in-pixel transistors.
Temperature differences across the array (Shading)
• White spots (pixels with very high leakage)

No of pixels
• Defect pixels

FPN (fixed pattern noise) for illuminated images


= PRNU (Photo response Non-Uniformity) σ

• White spots or defect pixels


• Deviations in signal darl level Outpur value
Constant with constant exposure
(Pixel variations)
Gain deviations in read out chain.
• Shading
Optical shading
Angle of incoming light (chief ray angel)
Gradient in capacitor values (determines the gain)
• Crosstalk
Light that hits neighbouring pixel
Charge carrier that diffuse from neighbouring pixel.

AO 10V 5.3
Noise INF 5440 - CMOS Image Sensors

Example on Shading (exaggerated)

Ref: Rick Baer (Aptina)

AO 10V 5.4
Noise INF 5440 - CMOS Image Sensors

TEMPORARY NOISE

AO 10V 5.5
Noise INF 5440 - CMOS Image Sensors

Photon noise
• m individual, independent and random events
• Average rate (number of events per time): n probability of n events within a given time interval.
Poisson distribution:
n
( n ) ( –n )
P ( n ) = ---------- e
n! (5.1)

Mean value of the m independent events:


nm = ∑ nm P ( n ) (m=1, 2, 3,...) (5.2)
n

It can be shown that in Poisson distributions, the variance is given by:


2 2 2
σ = n – (n) = n (5.3)

Mean value of number of photons that hit the sensor is n.


The rms noise (variance in number of photons) is the square root of the mean value.

[ref.: Ziel/King]

AO 10V 5.6
Noise INF 5440 - CMOS Image Sensors

Shot Noise in the Diode

Variation in number of electrons emitted over a


PN-junction (Schottky’s theorem).
Gives Poisson distribution i2

Determines the single event as q and qn= I (current).


For frequencies much lower than the transit time, the
spectral density becomes (energy densisty):

S i ( 0 ) = 2qI

Z e2 i2|Z|2
An average current I in a narrow frequency band ∆f i2 =

is equivalent with a noise current generator (2qI∆f)1/2


in parallel with the diode.

2
i = 2qI∆f (5.4)
Z

e2 = i2|Z|2

Ref.: Ziel

AO 10V 5.7
Noise INF 5440 - CMOS Image Sensors

Thermal noise
The random variable is a sum of a large number of independent statistic variables (pulses).
The pulse height X (discrete values) is normal distributed:
2 2
1 – ( X – X ) ⁄ 2σ
P ( X ) = ----------------- e
2
2πσ
In a continuous probability density function:
2 2
1 – ( x – x ) ⁄ 2σ
dP ( x ) = ----------------- e dx (5.5)
2
2πσ
The central limit theorem, that is when n is large: Independent variables: X1, X2, X3,..., Xn
• Equal probability density function
• Equal expectation value
• Equal variance:
n
• Y = ∑ Xi normal distributed for large n
i=1

• Y = nX 1

2 2
• σ y = nσ 1

AO 10V 5.8
Noise INF 5440 - CMOS Image Sensors

Thermal noise in resistance


The carriers thermal movements results in voltage pulses at the terminals of the resistor. Each pulse is short and has equal probability.
Based on the central limit theorem, normal distribution is considered. Because all single pulse has equal probability, all frequencies
has equal probability as well. Therefore we get a flat frequency spectrum SH(f) = SH(0), white noise.
The average free energy in a system in equilibrium is 3/2 kT (k=Boltzmanns constant, T=absolute temperature).
The average free energy of the current in the loop (one-dimensional system) is kT/2.
L
--- i = kT
2
------- (5.6)
2 2
di
Voltage around the loop: H ( t ) = Ri + L ----- written as fourier series:
dt

jω n t jω n t di jω n t
H(t) = ∑ αne i(t) = R ∑ βn e ----- = L
dt ∑ jω n β e
n
i(t)
n n n R
1
α n = ( R + jω n L )β n β n = α n ---------------------------- System function: 1/(R+jωnL) L
( R + jω n L )

Integrating the density function S over all frequencies to get the power value of the current: H(t)
∞ ∞ SH ( 0 ) ∞
2 SH ( 0 ) 2πL SH ( 0 )
∫ 2 ∫
i = S i ( f ) df = --------------------------------------
R + ( 2πf n ) L
2 2
- df = --------------- arc tan ---------- f
2πRL R n
0
= ---------------
4RL
0 0
Replacing i2 solved form (5.6) yields:(5.6)
2
S H ( 0 )∆f = v R = 4kTR∆f (5.7)
The Langevin Method
Ref.: Ziel

AO 10V 5.9
Noise INF 5440 - CMOS Image Sensors

Thermal Noise across a Capacitor


Average free energy for the circuit in equilibrium:
1 2 1
--- Cv = --- kT
2 2
The circuit has the transfer function: i(t)
vC ( f ) 1
-------------
- = ----------------------------- R
vH ( f ) 1 + j2πfRC C v

Power value:
H(t)
2 1
v C ( f ) = S H ( f ) -------------------------------------- (5.8)
2 2 2 2
1 + 4π f R C

The voltage source is a resistor with white noise SH(f) = SH(0).


Integrating over all frequencies with:
∞ ∞
2 1 SH ( 0 ) SH ( 0 )

v C = S H ( 0 ) -------------------------------------- df = --------------- arc tan ( 2πRCf ) = --------------
2 2 2 2 2πRC
0
4RC
- (5.9)
0 1 + 4π f R C

where SH(0) = 4kTR, such that:


2 4kTR kT
v C = -------------- = -------
4RC C
(5.10)
2 2 2 Ref.: Ziel
q C = v C C = kTC

AO 10V 5.10
Noise INF 5440 - CMOS Image Sensors

Noise in MOS-transistors
Two components
• Thermal, white, i2
2
in ∆f
carriers in thermal/statistic movements. ID
• 1/f
Frequency dependent.
Decreases with frequency as 1/f α
α ~1
1/f
Thermal
Transistor in saturation Thermal
2 2
∆i t = 4kT --- g m ∆f (5.11)
3
2
2 ∆i t 2 1 log f
∆v g = -------- = 4kT --- ------- ∆f (α =2/3) (5.12)
2 3 gm
gm
vn
Transistor in linear operation region in
2 1
∆v t = 4kT -------- ∆f (5.13)
g ds
2
2 ∆v t 2 2 (5.14)
∆i t = --------- = v t g ds = 4kTg ds ∆f
2
R
AO 10V 5.11
Noise INF 5440 - CMOS Image Sensors

1/f Noise
• Fluctuation of free carrier density in the channel due to “traps” in the interface Si/SiO2 and the close oxide.
• Fluctuation in the carrier mobility due to interaction with fonons and free path length.
• It can be shown (Ziel) that the statistic interaction has a wide range of time constants which gives
the 1/f spectrum.
Empiric expression for 1/f noise (ref. Tsividis):

2
2 Kg m 1
∆i f = ------------------- --- ∆f (5.15)
C ox WL f

2 Kf 1
∆v f = ------------------- --- ∆f (5.16)
C ox WL f

K is a technology dependent constant.

AF
2 ( KF )I D 1
SPICE model: i f = ---------------------- --- where KF and AF (~1) is SPICE parameters
C ox L f
2

AO 10V 5.12
Noise INF 5440 - CMOS Image Sensors

Random Telegraph Signals (RTS)


As MOS devices becomes smaller, capture of single
charge carriers becomes more dominant than
capturing av many carriers that creates the typical
1/f noise.
According to the theory, the mechanism similar, but
capture and release of single carriers make the
drain current to flip between 2 distinct values (in
addition to the thermal noise).

Images:
J.S. Kolhatkar et al., Separation of Random Telegraph Signals from
1/f Noise in MOSFETs under Constant and Switched Bias
Conditions

AO 10V 5.13
Noise INF 5440 - CMOS Image Sensors

Noise during sampling Th


R
Ts
vi vc
C

Th Ts
Tc
1/TC
Vi Vc
G(ω) H(ω,T)
When the switch (T) is conducting:
Fourier transform of the transfer function
1 1
G ( ω ) = ------------------------ = ------------------ (5.17)
1 + jωRC 1 + jωτ

Fourier transform of the sampled signal (from signal processing theory): Zero order hold funtion:
1 2πk
X ( jω ) = ------ V i ( jω )G ( jω )  jω a + j ---------- [ref. Oppenheim]
∑ (5.18)
Tc  Tc 
k

Vi(jω)*G(jω) corresponds to the noise described in (5.8).

Output form zero order hold-element: Ref.: Unbehauen/Vittoz


xf ( t ) = ∑ X ( nTh ) [ u ( t – nTh ) – u ( t – ( n + 1 )Th ) ]
n

AO 10V 5.14
Noise INF 5440 - CMOS Image Sensors

The LaPlace Transform


– snT h – s ( n + 1 )T h – snT h
e e e – sT h
Xf ( s ) = ∑ X n ( nT h ) --------------- – ---------------------------- =
s s ∑ X n ( nT h ) --------------- ( 1 – e
s
)
n n
– sT h – sT h
1–e – snT h 1–e Properties of
= ----------------------
s ∑ X n ( nT h ) [ e ] = ---------------------- X n ( s )
s the Laplace transform:
n 1
L [ u ( t ) ] = ---
Transfer function of the hold element: s
– sT h 1 – sT
Xn ( s ) 1–e L [ u ( t – T ) ] = --- e
H h ( s ) = -------------- = ---------------------- s
Xf ( s ) s

In the fourier version s=jω


– jωT h jωT h ⁄ 2 – jωT h ⁄ 2 T
1–e e –e h
H h ( jω ) = H h ( s ) = ------------------------- = ----------------------------------------------- ------
s = jω jω jωT h ⁄ 2 2
jωe ( Th ⁄ 2 ) Transform of a
sampeled sequense:
F∗ ( s ) = L [ f∗ ( t ) ]
sin ( ωT h ⁄ 2 ) – jωTh ⁄ 2
H h ( jω ) = T h ------------------------------- e (5.19) ∞
ωT h ⁄ 2 – snT
= ∑ f ( nT )e
n = –∞
Frequency response to the signal from the S/H system

V c ( jω ) = X ( jω )H h ( jω ) (5.18) (5.20)

AO 10V 5.15
Noise INF 5440 - CMOS Image Sensors

Consider the base component in the Fourier transform X(jω)


of the sampeled signal.

2 1 T h sin ( ωT h ⁄ 2 ) 2
Vc ( ω ) = 4kTR ------------------------------ ------ ------------------------------- (5.21)
1 + ω R C T c ωT h ⁄ 2
2 2 2

[Amplitude]2 High frequency region of the noise spectrum is folded down during
the sampling process.
R is the switch’s on resistance, C is the capacity of the sampling capacitor.
Integrating all frequences and with Th~Tc gives total noise power
equal to the total power before sampling:
2 kT
V c = ------- (5.22)
C

[Vc(ω)]2

[Vc(f)]2

1/Th 2/Th 3/Th f

AO 10V 5.16
Noise INF 5440 - CMOS Image Sensors

Correlated double sampling


High pass function - attenuates low frequency noise.
Sample function where the 2nd sample is subtracted form the 1st sample:

x ( n, τ d ) = x ( n ) [ δ ( t – nT h ) – δ ( t – nT h – τ d ) ] (5.23)

Simplify by regarding the double sampler only:


pδ ( t ) = δ ( t ) – δ ( t – τd ) (5.24)

F[δ(t)] = 1
Fourier transform: – jωT
jωτ d ⁄ 2 – j ωτ d ⁄ 2 F[δ(t – T)] = e
– jωτ d e –e ωτ – jωτ d ⁄ 2
P δ ( jω ) = 1 – e = ---------------------------------------------- = 2j sin  ------- e (5.25)
jωτ d ⁄ 2  2
e

ωτ
P δ ( jω ) = 2 sin  ------- (5.26)
 2

For ωτ << 1:
P δ ( jω ) = ωτ d (5.27)

Ref.: Unbehauen/Oppenheim

AO 10V 5.17
Noise INF 5440 - CMOS Image Sensors

Frequency response for double sampling and hold element

10 ωτ
2 sin  -------
 2
Amplitude [dB]
0 Combined

-10
T h sin ( ωT h ⁄ 2 )
------ -------------------------------
T c ωT h ⁄ 2
-20

-30

-40

-50

-60

-70
1 2 3 4 5
10 10 10 10 10 f [Hz]

AO 10V 5.18
Noise INF 5440 - CMOS Image Sensors

Reset Noise
At the end of the reset period (hard reset) the reset transistor
operates similar to a resistor.
Ron
It is a R-C circuit that gives a total noise of kT/C
Double sampling reduces this noise significantly. 4kTRon

Cfd
Sampling noise
2 kT
Noise current from the source follower transistor, current v rst = --------
C fd
source and noise current from the amplifier is added to the
noise generated by the switch’s on resistance.
Csmpl
Every noise current with a given spectral distribution can be isf ia,out
written as a noise voltage with its spectral distribution in
series with an output resistance. This voltage is shaped in
the sampling process as shown on the previous pages.
ics

Read-out noise
isfRo,sf Ro,sf
The voltage on the sampling capacitor is read by the
Csmpl ia,outRo,a
amplifier. Ro,a

The amplifier noise is added to the reset and sampling noise,


which again is shaped by the ADC sampler. Ro,cs
icsRo,cs

AO 10V 5.19
Noise INF 5440 - CMOS Image Sensors

Noise considerations in the Source Follower


Thermal noise
Noise bandwidth: Assuming rectangular shaped frequency spectrum with all its power inside the rectangle.
Normalized white noise: |H(jω)|=1
∞ 2 ∞
H ( jω ) 1
1
∆f = ------------------------------- ∫
----------------------------- dω
2 = 2π ∫ (---------------------------------
2
- df
=
π
-----------
H ( jω ) max 0 ( ωRC ) + 1 2πRCf ) + 1
2 2RC
0
H ( jω ) = 1

Equivalent thermal noise voltage at the input:

M1 2 2 1
∆v th = 4kT --- ------- ∆f
3 gm

Vgs Assuming a large part of the noise as uncorrelated and


band limited by SF’s output resistance 1/gm and
Id load capacitance Csh = Cshr = Cshs
SHS SHR
M2
Cshs Cshr 2  2 1 gm  8π kT
Vb2 v th, shr ≈ 2  4kT --- ------- ------------ π = ------ --------- (5.28)
 3 g m 2C sh  3 C sh

AO 10V 5.20
Noise INF 5440 - CMOS Image Sensors

Flicker noise

2 Kf 1
∆v f = - --- ∆f
------------------
C ox WL f

Flicker noise has a large low frequency part, which is attenuated by double sampling:
Total noise in differential signal (equivalent input noise at the SF input):

2 Kf 1 1

v 1 ⁄ f = 2 ------------------- --- ---------------------------------- ( 1 – cos ( 2πfτ ) ) df
C ox WL f ( 2πfRC ) 2 + 1
(5.29)

The last term is from (5.26)

2 ωτ 2 ωτ 2 1 1
P δ ( jω ) = 2 sin ------- = 4  1 –  cos -------  = 4  1 – --- – --- cos ( ωτ ) = 2 ( 1 – cos ( ωτ ) )
2   2    2 2 

AO 10V 5.21
Noise INF 5440 - CMOS Image Sensors

RTS noise
Combined with the Correlated Double sampling, three distinct levels are created

Images: Xinyang Wang et al: Random Telegraph Signal in CMOS Image Sensor Pixels.

AO 10V 5.22
Noise INF 5440 - CMOS Image Sensors

Noise sources - summary

in2

in2

2qIph
kT/C
kT/C
ADC
vn2
quantization
noise
vn2
kT/C

in2

Comment: Low frequency noise is filtered by double sampling

AO 10V 5.23
Noise INF 5440 - CMOS Image Sensors

Equivalent noise reference

Noise voltage referred to the output

2 2 2 2
v n, out = ( A v CG ) qσ n, pix + v n, signchain [V] (5.30)

Noise charge, in number of electrons, referred to the input


2
2 v n, signchain 2 [e-] (5.31)
σ n, in = ------------------------------ + σ n, pix
2
q ( A v CG )

Can be given as equivalent luminance (lux) by including QE and optical loss.

AO 10V 5.24
Noise INF 5440 - CMOS Image Sensors

Definitions
Dynamic range - DR
The ratio largest possible signal to the signal that is just at the noise level in one image: Intra scene dynamic range.
 N sat 
DR [ dB ] = 20 log  ------------- (5.32)
 n read

Signal to Noise Ratio - SNR


The ratio signal, independent of the level, and the noise.
The Poisson distributed photon noise, σ = Nsig , is level dependent.

 N sig
SNR [ dB ] = 20 log  ----------- (5.33)
 n tot 

where ntot is read noise plus photon noise.


At large signal levels, the photon noise is dominating.
 N sig   N sig 
SNR high [ dB ] = 20 log  --------------- = 20 log  --------------- = 20 log N sig (5.34)
 n foton  N sig

and maximum SNR is achieved close to saturation:

SNR max [ dB ] = 20 log N sat (5.35)

AO 10V 5.25
Noise INF 5440 - CMOS Image Sensors

AO 10V 5.26
Noise INF 5440 - CMOS Image Sensors

Reference

Ziel
Noise in Solid State Devices and Circuits
Aldert van der Ziel
John Wiley & sons (1986)

King
Electrical Noise
Robert King
Chapman and Hall Ltd.

Vittoz
Dynamic Analog Techniques by Eric A. Vittoz
Design of MOS VLSI Circuits for Telecommunications
Editors: Y. Tsividis / P.Antognetti (Editors)
Prentice-Hall (1985)

Oppenheim
Digital Signal Processing
Alan V. Oppenheim / Ronald W. Schafer
Prentice-Hall

Nakamura
Image Sensors and Signal Processing for Digital Still Cameras,
edited by Junich Nakamura
Taylor & Francis

AO 10V 5.27
Noise INF 5440 - CMOS Image Sensors

AO 10V 5.28
Noise INF 5440 - CMOS Image Sensors

Example: Read-out chain

• 2 M pixler
• 15 Frames per second
• Exposure time 10 ms
To be discussed:
• The requirements for the speed in
Source Follower,
in gain amplifier,
and ADC
• Required gain settings
• Required ADC resoution

AO 10V 5.29

Das könnte Ihnen auch gefallen