noise at the system levelkilopixel.net/publications/isscc 2007 - rlb v4.pdf · 2020. 10. 18. ·...
TRANSCRIPT
-
ISSCC 2007 Forum:Noise in Imaging Systems
February 15, 2007
Noise at the System LevelRick Baer
-
21/5/2007
Introduction
-
3
Micron Confidential
1/5/2007
The systems view
Noise sources
Noise modifiers
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
4
Micron Confidential
1/5/2007
Major CCD and CMOS noise sources
• BothFlareVignettingPRNU (pixel lithography)Photon shot noiseRead noise (thermal, 1/F, kT/C)Row reference
• CCDColumn dark current variationSmearBlooming
• CMOSColumn gain variationPGA / ADC gain and offset variation
-
5
Micron Confidential
1/5/2007
Major noise modifiers
Increases noise:
• White balance
• Color correction
• Sharpening
• Tone mapping
Decreases noise:
• FPN correction
• Noise reduction filtering
• compression
Diffuses noise:
• Pixel reconstruction
• Color space conversion
-
61/5/2007
Overview
• There are many sources of noise and interference in imaging systems
• Image processing has a big effect on noise at the systems level
• Noise measurement is complicated by the (unknown) effects of image processing
-
71/5/2007
Noise from the perspective of image processing
-
81/5/2007
Pixel / circuit noise sources(reviewed in other ISSCC noise forum presentations)
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
91/5/2007
Exposure control
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
10
Micron Confidential
1/5/2007
Simplified noise / exposure control model
X ADC
Texp Analog gain
Dark currentnoise
+ +Xsignal
Read noise (N1)
+
ADC input, quantization noise (N2)
Exposure controls
-
11
Micron Confidential
1/5/2007
SNR variation
2
2N22
N1 G/S σσ +=S/N
Gain
S/N
mi ma
The signal to noise ratio increases with gain until it is limited by noise source N1.
-
12
Micron Confidential
1/5/2007
DR variation
2N2
2N1
2G/552 σσ +=k
DR
Gain
DR
mi ma
The dynamic range decreases with gain.
-
13
Micron Confidential
1/5/2007
Exposure
Short exposure(noise, quantization)
Long exposure(clipping)
Ideal exposure
-
14
Micron Confidential
1/5/2007
Flicker suppression
desk lamp
0.6000
0.8000
1.0000
1.2000
1.4000
1.6000
1.8000
0 2 4 6 8 10 12 14 16 18
time [msec]
rela
tive
inte
nsity
60W tungfluor ring
Restrict exposure and frame periods to multiples of the flicker period (e.g. 1/120 Hz)
-
15
Micron Confidential
1/5/2007
What’s wrong with this picture?
-
16
Micron Confidential
1/5/2007
Importance of exposure control
• Underexposure leads to poor SNR and possibly posterization
• Underexposure may lead to poor color reproduction (due to pixel non-linearity)
• Overexposure leads to clipping
• Incorrect exposure timing may cause flicker to appear
-
171/5/2007
Offset / gain correction
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
18
Micron Confidential
1/5/2007
Sources of offset FPN
• Row FPN Optical black clamp circuit (CCD & CMOS)Power supply noise (CMOS)
• Column FPNDark current (CCD)
Column offset (CMOS)
• Random FPNDark current (CCD & CMOS)
-
19
Micron Confidential
1/5/2007
CCD and CMOS readout timing
row 1
row 2
row n
row 1
row 2
row nreset
integrate
Transfer (readout)digitize
time
CDS row 1row 2
row n
CCD
“SCDS”4T RS CMOS
3T RS CMOS
Temporal separation introduces susceptibility to dark current, power supply noise
-
20
Micron Confidential
1/5/2007
Visibility of row/column FPN
random / row ratio
21 5 10
Row noise is still visible, even when it is buried in 5x random noise
-
21
Micron Confidential
1/5/2007
CCD optical black reference pixels
leadinghorizontal
OB trailinghorizontal
OB
leadingvertical
OB
trailingvertical
OB
Sony ICX085
-
22
Micron Confidential
1/5/2007
Horizontal optical black signal vs. row number
average of 40 columns
-
23
Micron Confidential
1/5/2007
CCD row FPN offset correction (OB clamp)
OB accumulator
+−
+
Pixel data
Pixel ID
Methods:1. Accumulate leading/trailing OB pixels (analog / digital)2. Multi-row running average (digital)3. Exclude outlying pixels (digital)
-
24
Micron Confidential
1/5/2007
CCD column FPN (vertical register dark current)
Generally buried in read noise: visible at high temperatures
-
25
Micron Confidential
1/5/2007
CMOS row FPN offset correction
S/H
S/H
Σ
+
S/H
Σ
+
S/H
Σ
+
S/H
Σ
+
average
- - - -
Pixel circuit
Column bufferOptical black
-
26
Micron Confidential
1/5/2007
CMOS row offset features
• Analog or digital offset subtraction• Provides power supply noise rejection (for 3T CMOS)
• Dark current clamping isn’t required
• Dark current can introduce row noise!
-
27
Micron Confidential
1/5/2007
Sources of gain FPN
• PRNU (random)• Column gain variation (column)
• ADC/PGA gain variation (block / color plane)
-
28
Micron Confidential
1/5/2007
Gain FPN (multiplicative noise)
Column gain
ADC/PGA gain (four blocks)
PRNU
Assign each color plane to a separate ADC/PGA to avoid visible gain errors.
-
29
Micron Confidential
1/5/2007
Gain FPN correction
Inverse flat-field image
XData in
address
Data out
1/(column gain)
XData in
address
Data outData in AWB
Data out
PRNU
column
PGA/ADC
Digital correction and memory is required!
-
30
Micron Confidential
1/5/2007
Effect of offset / gain correction
• FPN that would otherwise be intolerable may be almost completely eliminated
• Row FPN correction circuits may introduce their own FPN
• Spatially random offset / gain correction (dark current / PRNU) requires large memories (and large digital multipliers)
• Offset correction may suppress power supply noise
-
311/5/2007
Defect correction
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
32
Micron Confidential
1/5/2007
Defect correction
• Static defect tables• Dark current – works best on extraordinary pixels,
not on distribution tail
• On-the-fly correctionThreshold problems (video mode)
• Effect of correction (interpolation -> spatial correlation)
-
33
Micron Confidential
1/5/2007
Defect correction
Defect location
table
bypass / interpolate
Data in
addressaddress
Data out
Row buffers interpolator
exception detectorData in
Data out
Static defect correction
Dynamic defect correction
Row buffers interpolator
-
34
Micron Confidential
1/5/2007
Defect characteristics
White defects• Caused by dark current “tail”, impurities, dislocations• Temperature dependent• Generally single pixels• Correction works best on isolated “hot” pixels
Black defects• Caused by occlusions• Generally clusters of pixels
Dark current
Coun
t
Dark current
Coun
t
?
-
35
Micron Confidential
1/5/2007
Effects of defects / defect correction
• Defective pixels contribute additive or multiplicative noise
• Defect correction replaces bad pixels with interpolates of neighboring values, results in resolution loss.
• Resolution loss isn’t visible unless the number of defective pixels is large (> 0.1%).
• On-the-fly defect correction may introduce temporal noise (blinking pixels)
-
361/5/2007
Flare Compensation (ABL)
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
37
Micron Confidential
1/5/2007
Sources of flare
Reflections from:•Lens surfaces•Lens barrel•Aperature•IR filter•Cover glass•Sensor•Bond wires and pads
-
38
Micron Confidential
1/5/2007
Ghost images
-
39
Micron Confidential
1/5/2007
Non-uniform flare
-
40
Micron Confidential
1/5/2007
Reference column leakage
-
41
Micron Confidential
1/5/2007
Uniform flare compensation
10% flare
0% flare
-10% flare
-
42
Micron Confidential
1/5/2007
Effects of flare
• Uniform flare adds signal-dependent offset (noise)• Uniform flare can be subtracted. Over-subtraction
reduces color accuracy.
• Specular sources may produce ghost images which can’t be removed.
• Black row effects may appear because of light leaking into row reference pixels.
-
431/5/2007
Shading correction
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
44
Micron Confidential
1/5/2007
Vignetting and color shading
Lens vignetting: color planes share a common
center
Color shading: color planes have different centers (because
of pixel asymmetry)
Sources of shading•Lens vignetting•IR filter CRA dependence•Pixel angular response / asymmetry
-
45
Micron Confidential
1/5/2007
Shading correction
Uncorrected image
Corrected image
X
X
X
X
shading correction parameters
(from camera characterization)
Shading model
Approximate inverse flat-
field
Approximate inverse flat-
field
-
46
Micron Confidential
1/5/2007
Illuminant-dependent color shading
Shading correction optimized for CWF
D65 - daylight CWF - fluorescent “A” - tungsten
(color saturation increased to enhance visibility)
-
47
Micron Confidential
1/5/2007
Effects of vignetting / pixel angular response
• Reduced signal / SNR in corners• Color shading
• Gr/Gb channel imbalance
-
481/5/2007
Pixel reconstruction
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
49
Micron Confidential
1/5/2007
Pixel reconstruction
• Undersampling results in artifacts• Demosaic uses interpolation -> results in spatial
correlation of noise.
-
50
Micron Confidential
1/5/2007
Color artifacts caused by undersampling
-
51
Micron Confidential
1/5/2007
Noise correlation
G G G
G
GGG
G G
B B
B B
B BG G
GGR R R
R R R
Noise diffuses to adjacent pixels
-
52
Micron Confidential
1/5/2007
Effects of pixel reconstruction
• Reconstruction may introduce color artifact noise• Reconstruction diffuses noise to adjacent pixels
-
531/5/2007
White balance
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
54
Micron Confidential
1/5/2007
Effect of color temperature on RGB balance
red
green
blue
elec
tron
s/lu
x
Color temperature [K]
Sony CCD QE curves and CM-500 IR filter
-
55
Micron Confidential
1/5/2007
White balance operation
X
X
X
Σ
red stats
Σ Σ
green stats
blue stats
R
G
B
Scene rule base (e.g.
gray world)
-
56
Micron Confidential
1/5/2007
Effect of white balance
• Neutral colors balance to gray• Amplification of weak color channels increases noise
• Relative color strength depends on spectral response, illuminant color
-
571/5/2007
Color correction
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
58
Micron Confidential
1/5/2007
Color Processing Decreases the SNR
CA ⋅=BGR A transformation matrix is required to convert from the sensor color
space to the output color space
Consider a two-color world:
BRCBRC
2
1
+=+=
αα
Transform to color: Transform to monochrome:
212
221
1CCB
1CCR
αα
αα
−−
=−−
= ( )2121 CCX +=
( )22
221
21
NNSSNS R
αα
+
−= ( ) 2
221
21
NNSS
NS X
+
+=
• Color correction decreases the SNR• The greater the crosstalk, the greater the noise amplification
α = pixel crosstalk factor
sensoroutput
-
59
Micron Confidential
1/5/2007
Effect of color correction
• Color fidelity is improved
• Noise is significantly amplified by off-diagonal matrix terms
• Off-diagonal matrix terms depend on spectral response and pixel-to-pixel crosstalk
-
601/5/2007
Sharpening
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
61
Micron Confidential
1/5/2007
Sharpening filters amplify noise
original sharpen
original + noise
-
62
Micron Confidential
1/5/2007
Effect of sharpening
• The noise has more high-frequency spectral content than the image
• Sharpening filters amplify high-frequencies and therefore amplify noise
-
631/5/2007
Noise reduction filtering
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
64
Micron Confidential
1/5/2007
Noise filtering(topic of Dr. Pizurica’s ISSCC presentation)
• Temporal filtering• Spatial filtering
• Adaptive spatial filtering
-
65
Micron Confidential
1/5/2007
What distinguishes signal from noise?
• The signal is stationary (… but so is FPN)• The spatial frequency response of the signal is
limited by the system MTF
• The color gamut of the signal is limited by the surfaces / illuminants observed in the real world
-
66
Micron Confidential
1/5/2007
Effect of noise reduction
• Noise is presumably reduced• Sharpness may be lost
• Texture detail may be lost
-
671/5/2007
Gamma correction and tone mapping
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
68
Micron Confidential
1/5/2007
Gamma correction and tone
• Gamma correction – compensates for display device• Global tone correction – produces pleasing tone
reproduction
• Local tone correction (e.g. Retinex) – for high DNR images
-
69
Micron Confidential
1/5/2007
Effect of tone correction on SNR
input
outp
ut
Gamma / tone function
Noise amplification
= slope of tangent line
signal amplification
= slope of line from origin
original
after gamma correction
The tone correction function must be reversed to make meaningful
noise measurements
-
701/5/2007
Color space conversion
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
71
Micron Confidential
1/5/2007
Color space conversion
R G B R G B Y u Y v Y u Y u
2 pixels 2 pixels
conversion
-
72
Micron Confidential
1/5/2007
Chrominance filtering
LPF
luminance
chrominance
-
73
Micron Confidential
1/5/2007
Effects of color space conversion
• Chrominance noise may be reduced• Noise may be spatially diffused
-
741/5/2007
Compression
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
-
75
Micron Confidential
1/5/2007
Compression
•Noise reduced•Severe block artifacts
•Noise significantly reduced•Loss of texture
JPEG 2000compression
JPEG compression
27x 27x
-
76
Micron Confidential
1/5/2007
Effects of compression
• Noise is reduced (similar to noise reduction processing)
• New sources of noise (block artifacts, quantization) may be added
-
771/5/2007
Measurement
-
78
Micron Confidential
1/5/2007
Measurement challenges
• Interpretation of luminance versus chrominance noise
• Dependence on viewing conditions (Johnson & Fairchile EI2004)
High frequency chromatic noise doesn’t effect quality• Nonlinear processing (un-map gamma)• Quantization limits on measurement• Noise decomposition (temporal, fixed, etc.)• (other practical stuff like shading suppression)• Compression / adaptive noise filtering
-
79
Micron Confidential
1/5/2007
ISO 14524 OECF measurement
Diffusing screen
Extended source
OECF test chart
System under test
Required to un-map gamma & tone correction functions
-
80
Micron Confidential
1/5/2007
ISO 9358:1994 veiling glare measurement
• Integral method (suitable for uniformly radiant scenes)
• Analytical Method (suitable for intense isolated sources)
Diffusing screen
Extended source
Black area
System under test
System under test
source
collimator
-
81
Micron Confidential
1/5/2007
ISO 15739 noise measurement
Diffusing screen
Extended source
OECF test chart
System under test
OECF measurement
patches
Noise measurement
patches
High frequency
measurement patches
-
82
Micron Confidential
1/5/2007
ISO 15739 noise component analysis (1)
++ -
FPN image
Temporal noise
images
(+ σ temporal)
>= 8 images
-
83
Micron Confidential
1/5/2007
ISO 15739 noise component analysis (2)
+++-
Spatially-random noise
image
FPN or temporal
noise image-
Row noise vector
Column noise vector
(+ σ random)
(+ σ random)
-
84
Micron Confidential
1/5/2007
Utility of noise decomposition
-
85
Micron Confidential
1/5/2007
Quantization limits
Noise below ½ LSB can’t be measured
-
861/5/2007
Conclusions
-
1/5/2007 87
871/5/2007
Image processing modifies noise …
pixel array
analog / ADC
offset FPN
correct
gain FPN
correct
defect correct
flare correct
shading correct
Exposure control
scene intelligence
white balance
color correct sharpen
noise reduce
gamma & tonedemosaic
compresscolor space output
Cancels / reduces noise
Generally increases noiseIncreases noise
-
1/5/2007 88
881/5/2007
… and makes it difficult to measure!
• Opto-electronic conversion function• Luminance vs. chrominance
• Adaptive processing
• Compression artifacts
The ISO 12232 digital photography speed measurement standard has not been widely adopted because of the difficulty of noise measurement.
-
1/5/2007 89
891/5/2007
Noise is one dimension of image quality …
Noise
Texture detail
Color saturation Resolution
Tone reproduction
Artifacts
-
1/5/2007 90
901/5/2007
… but they all matter!
Noise
Texture detail
Color saturation Resolution
Tone reproduction
Artifacts
-
91
Micron Confidential
1/5/2007
-
92
Micron Confidential
1/5/2007
References
• J. Nakamura, Image Sensors and Signal Processing for Digital Still Cameras, CRC Press, New York, (2005).
• S. Battiato, M. Mancuso, “An Introduction to the Digital Still Camera Technology”, ST Journal of System Research - Special Issue on Image Processing for Digital Still Camera, Vol. 2, No.2, (12/2001).
• G. Williams Jr., Ed., Digital Solid State Cameras: Designs and Applications, Proc. SPIE, Vol. 3302, (1998).
• N. Sampat, J. DiCarlo, R. Motta, Eds., Digital Photography, Proc. SPIE, Vol. 5678, (2005).
• N. Sampat, J. DiCarlo, R. Martin, Eds., Digital Photography II, Proc. SPIE, Vol. 5678, (2006).
• R. Martin, J. DiCarlo, N. Sampat, Eds., Digital Photography III, Proc. SPIE, Vol. 5678, (2007).
• J. Lopez-Alonso, R. Gonzalez-Moreno, J. Alda, “Noise in imaging systems: fixed pattern noise, electronic, and interference noise”, Proc. SPIE, Vol. 5468, p. 399-407, (5/2004).
• ISO 9358 – Optics and optical instruments – Veiling glare of image-forming systems – Definitions and method for measurement.
• ISO 12232 – Photography – Digital still cameras – Determination of exposure index, ISO speed ratings, standard output sensitivity, and recommended exposure index.
• ISO 14524 – Photography – Electronic still-picture cameras – Methods for measuring opto-electronic conversion functions (OECFs).
• ISO 15739 – Photography – Electronic still-picture imaging – Noise measurements.
Noise at the System LevelIntroductionThe systems viewMajor CCD and CMOS noise sourcesMajor noise modifiersOverviewPixel / circuit noise sourcesExposure controlSimplified noise / exposure control modelSNR variationDR variationExposureFlicker suppressionWhat’s wrong with this picture?Importance of exposure controlOffset / gain correctionSources of offset FPNCCD and CMOS readout timingVisibility of row/column FPNCCD optical black reference pixelsHorizontal optical black signal vs. row numberCCD row FPN offset correction (OB clamp)CCD column FPN (vertical register dark current)CMOS row FPN offset correctionCMOS row offset featuresSources of gain FPNGain FPN (multiplicative noise)Gain FPN correctionEffect of offset / gain correctionDefect correctionDefect correctionDefect correctionDefect characteristicsEffects of defects / defect correctionFlare Compensation (ABL)Sources of flareGhost imagesNon-uniform flareReference column leakageUniform flare compensationEffects of flareShading correctionVignetting and color shadingShading correctionIlluminant-dependent color shadingEffects of vignetting / pixel angular responsePixel reconstructionPixel reconstructionColor artifacts caused by undersamplingNoise correlationEffects of pixel reconstructionWhite balanceEffect of color temperature on RGB balanceWhite balance operationEffect of white balanceColor correctionColor Processing Decreases the SNREffect of color correctionSharpeningSharpening filters amplify noiseEffect of sharpeningNoise reduction filteringNoise filtering�(topic of Dr. Pizurica’s ISSCC presentation)What distinguishes signal from noise?Effect of noise reductionGamma correction and tone mappingGamma correction and toneEffect of tone correction on SNRColor space conversionColor space conversionChrominance filteringEffects of color space conversionCompressionCompressionEffects of compressionMeasurement challengesISO 14524 OECF measurementISO 9358:1994 veiling glare measurementISO 15739 noise measurementISO 15739 noise component analysis (1)ISO 15739 noise component analysis (2)Utility of noise decompositionQuantization limitsImage processing modifies noise …… and makes it difficult to measure!Noise is one dimension of image quality …… but they all matter!References