IMAGE NOISE I
• APPLICATIONS
• Signal estimation in presence of noise
• Detecting known features in a noisy background
• Coherent (periodic) noise removal
ECE/OPTI533 Digital Image Processing class notes 238 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I
TYPES OF NOISE
• photoelectronic
• photon noise
• thermal noise • impulse
• salt noise
• pepper noise
• salt and pepper noise
• line drop • structured
• periodic, stationary
• periodic, nonstationary
• aperiodic
• detector striping
• detector banding
ECE/OPTI533 Digital Image Processing class notes 239 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I
Photoelectronic noise
• Photon noise
Photon arrival statistics
Low-light levels (nightime imaging, astronomy)
• Poisson density function
• Standard deviation = square root mean (signal-dependent)
High-light levels (daytime imaging)
• Poisson distribution —> Gaussian distribution
• Standard deviation = square root mean • Thermal noise
Electronic
White (flat power spectrum), Gaussian distributed, zero-mean (signal-independent)
ECE/OPTI533 Digital Image Processing class notes 240 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I • Photoelectronic noise model
Photon noise is signal-dependent
Thermal noise is signal-independent
One model for a combined noise field fh (m, n) is:
fh (m, n) = h P(m, n) fs(m, n) + h T(m, n)
where
h P(m, n) and h T(m, n) are independent white, zero-mean Gaussian noise fields
fs(m, n) is the noiseless signal (may not be measurable)
Note, h P(m, n) has unit standard deviation and is scaled by square root of signal
• Approximates photon noise component for large signals
ECE/OPTI533 Digital Image Processing class notes 241 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I • Noisy image model
f(m, n) = fs(m, n) + fh (m, n) = fs(m, n) + h P(m, n) fs(m, n) + h T(m, n)
additive signal-dependent and signal-independent random noise
• Note, this model may not apply in particular situations!
ECE/OPTI533 Digital Image Processing class notes 242 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Examples of simulated thermal noise for different noise standard deviations s h
5
20 10
ECE/OPTI533 Digital Image Processing class notes 243 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Examples of simulated photon + thermal noise for different standard deviations s h
5
10 20
ECE/OPTI533 Digital Image Processing class notes 244 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I IMPULSE NOISE pepper noise (0.05% and 2%)
• Data loss or saturation
• Definitions
• Salt noise: DN = maximum possible
• Pepper noise: DN = minimum possible
• Salt and pepper noise: mixture of salt and pepper noise
• Line drop: part or all of a line lost
ECE/OPTI533 Digital Image Processing class notes 245 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I
Line drop
ECE/OPTI533 Digital Image Processing class notes 246 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I STRUCTURED NOISE simulation example
Periodic, stationary
• Noise has fixed amplitude, frequency and phase
• Commonly caused by interference between electronic components
ECE/OPTI533 Digital Image Processing class notes 247 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Mars Mariner example - multiple frequencies (Rindfleish et al, 1971)
ECE/OPTI533 Digital Image Processing class notes 248 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Periodic, nonstationary
• noise parameters (amplitude, frequency, phase) vary across the image
• Intermittant interference between electronic components simulation example
ECE/OPTI533 Digital Image Processing class notes 249 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I
Mars Mariner 9 example - single frequency, variable amplitude (Chavez and Soderblum, 1975)
ECE/OPTI533 Digital Image Processing class notes 250 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Aperiodic
• JPEG noise
JPEG-compressed (low quality)
difference (noise)
ECE/OPTI533 Digital Image Processing class notes 251 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I • ADPCM (Adaptive Pulse Code Modulation) noise
• IKONOS 1-m panchromatic imagery
• Kodak proprietary compression algorithm
lake in Reid Park, Tucson DN 200-220 contrast-stretched
ECE/OPTI533 Digital Image Processing class notes 252 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Detector Striping example with 4 detectors
• Calibration differences among individual scanning detectors
detector 1 2 . scan j i . N detectors/ N scan 1 2 . i scan direction . reverses N
• For detector i:
DNi = gainiE + offseti
where E is the scanned optical image
ECE/OPTI533 Digital Image Processing class notes 253 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I Detector Banding
• Calibration changes from scan-to-scan (whiskbroom scanner)
detector 1 2 . scan j i . N detectors/scan N 1 2 . scan direction i reverses . N
• For detector i, scan j:
DNij = gainj(gainiE + offseti) + offsetj
where E is the scanned optical image irradiance (W-m-2)
• Changes in gainj or offsetj from scan-to-scan can be caused by detector saturation at one end of scan
ECE/OPTI533 Digital Image Processing class notes 254 Dr. Robert A. Schowengerdt 2003 IMAGE NOISE I
example Landsat Thematic Mapper (Schowengerdt, 1997) - 16 detectors/scan
original (San Francisco Bay) water mask
masked contrast- original stretched
ECE/OPTI533 Digital Image Processing class notes 255 Dr. Robert A. Schowengerdt 2003