Noise Characterization from NonRaw CCD Output

1
Noise Characterization from Non-Raw CCD Output

• Gregory Ng
• March 10, 2005

2
Goals

• Observe noise in consumer-level digital cameras
• What is ultimate effect on image?
• Is it unique to particular models?
• Characterize camera system noise as far as
  possible
• Define features needed to perform
  characterization tests

3
Overview

• Cameras
• Special Considerations
• Illustrations
• Noise Statistics
• Sources
• What to do about it?

4
Cameras
A70, S1 IS, A80(Canon)
Nikon D70
Canon S30
Quickcam Pro 4000
Nikon Coolpix 4300
5
Types of Noise

• Fixed-Pattern Noise
• Random Noise
• Dark Current
• Shot Noise
• Read Noise

S1 IS ISO 50, 1 sec
6
Special Considerations

• Non-Raw Data
• Nonlinear output
• White balancing
• JPEG compression

7
ISO Settings
ISO 50- Long exposure
ISO 100

• ISO 400
• Noise grain
• JPEG artifacts
• White balance

steves-digicams.com
8
Noise
- Mosaicing pattern affects noise spatial
characteristics- Amplifier gain affects
intensity
9
Confirming JPEG Noise

• Blocking Artifacts
• 2-D FFT
• Color Quantization
• Confirmed by S30 (uncompressed)

Simulated Image Noise
FFT of Simulated Noise
10
Noise Statistics S1 IS
Sample Variance (FSCS)

• Variance
• Var(R) 4.38
• Var(G) 2.80
• Var(B) 9.47
• Covariance
• Cov(R,G) 0.568
• Cov(R,B) 2.181
• Cov(G,B) 0.575

11
Noise Statistics S30
Sample Variance (FSCS)

• Variance
• Var(R) 1.37
• Var(G) 1.32
• Var(B) 4.00
• Covariance
• Cov(R,G) 0.19
• Cov(R,B) 0.55
• Cov(G,B) 0.82

12
Noise Statistics QC4000
Sample Variance (FSCS)

• Variance
• Var(R) 4.76
• Var(G) 4.31
• Var(B) 7.66
• Covariance
• Cov(R,G) 3.40
• Cov(R,B) 3.78
• Cov(G,B) 3.17
• Not JPEG artifacts (4x4 blocks)

13
Noise Statistics A70,A80
Sample Variance (FSCS)
A70
A80
14

• Radar plot indicates variance, not intensity of
  noise
• Bayer pattern explains high red, blue variance
• High Cov(R,B) explained by noise in green pixels

15
Signal-Level Dependence

• Test Procedure
• Aim camera at LCD
• Defocus to avoid pixel artifacts, Moire
• Manual exposure
• Monitor gamma irrelevant
• High contrast (5001)
• Obtain 20 sample images
• Crop to give flat histogram

16
Noise
Variance
SNR
17
Comments

• Noise spatially, chromatically similar across all
  tested cameras
• S30 noise levels exceptionally low at ISO 50
• Unexplained drop in noise variance at high signal
  levels

18
Noise Sources

• Shot noise (sshot sqrt(S))
• Amplifier noise
• Ignored
• Dark current noise
• Negligible energy 0.15 to 0.16 ADU/channel at
  1 exposure
• PRNU (not tested)
• Stuck pixel visually detected
• Other sources
• Reset noise
• Output amplifier noise
• Flicker noise (1/f noise)
• Electrical system noise (banding)

(blue channel)
Kodak Application Note MTD/PS-0233CCD Image
Sensor Noise Sources
19
What to do?

• Fixed-pattern noise -- easy
• Random noise
• Lower ISO
• Average multiple images
• Gaussian blur (a.k.a. resample)
• Median filter
• PSP Edge-preserving smooth

20
Removing Noise

• Neat Image

21
Neat Image
www.neatimage.com
22
Conclusions

• ISO measures exposure/noise tradeoff
• Additional steps needed to get CCD parameters
• Near photon-limit noise
• Postprocessing removes much noise
• Canon S1 IS noise is no different from others
• IS allows lower-light shooting