Practical Poissonian-Gaussian Noise Modeling and Fitting for Single-image Raw-data Alessandro Foi, Mejdi Trimeche, Vladimir Katkovnik, and Karen Egiazarian Department of Signal Processing, Tampere University of Technology
Noise Modeling for General Cases Noisy Images Variant Acquisition Devices Image Processing Output Images Noise Model It is hard to construct a general noise model!
Noise Modeling for Specific Case Noisy Images Image Processing Output Images Noise Model Foi, A., M. Trimeche, V. Katkovnik, and K. Egiazarian, “Practical Poissonian-Gaussian noise modeling and fitting for single-image raw- data”, IEEE Trans. Image Process., vol. 17, no. 10, pp , October 2008.
An Example of Application: Sharpness Enhancement Noisy Image Sharpness Enhancement Output Image Noise Model Original Image Enhanced Image with Noise Suppression
Application with Noise Model Spatial Index Intensity
Application with Noise Model Scene Radiance Image Sensor Camera Processing Output Image Camera System Denoising, Demosaicing, Deblurring, Compression Raw-Data Noise Modeling
Noise Model Scene Light Photon Sensor Electrical Devices Image Raw-data Poisson Noise Gaussian Noise Signal-dependentSignal-independent Camera Sensor
Noise Model
Poissonian-Gaussian Noise Poisson: Gaussian: Overall Variance of z: Overall Standard-deviation of z:
Noise Level Function Intensity Standard-deviation Intensity Standard-deviation a = , , , b = a = 0.4 2, b = , ,
Poissonian-Gaussian Noise
Raw-data Modeling Overall Standard-deviation of z: 1.Quantum Efficiency 2.Pedestal Parameter 3.Analog Gain Efficiency ↑ a↓ The percentage of photons hitting the photoreactive surface that will produce an electron
Noise Model Sensor Model Noise Model
Sensor Parameter and Noise Parameter Sensor Parameter: ISO-number Gain Temperature Shutter Time
Sensor Parameter and Noise Parameter
Two Stages of Noise Estimation Local estimation of multiple expectation/standard-deviation pairs. Global parametric model fitting. Intensity Standard-deviation
Two Stages of Noise Estimation Local Estimation Intensity Standard-deviation Global Parametric Model Fitting
Local Estimation Local Expectation/ Standard-deviation Pair Locally Smoothed Value Locally Detail Value
Wavelet Analysis Local Expectation/ Standard-deviation Pair Locally Smoothed Value Locally Detail Value Noise Component
Wavelet Analysis Noise Component For Smooth Region
Wavelet Analysis Noise Component For Smooth Region
Smooth Region Segmentation Smooth Region Segmentation Edge Detection Smoothing
Level Sets Segmentation Smooth Region Segmentation Smooth Region … Level Segmentation Intensity 01 …
Local Estimation of y i … Estimation
Local Estimation of σ i … This factor, which comes from the mean of the chi-distribution with n − 1 degrees of freedom, makes the estimate unbiased for normally and identically independently distributed (i.i.d.)
Global Fitting Intensity Standard-deviation
Clipping Effect Intensity Spatial Index 1 0 Intensity Spatial Index 1 0 Clipping from above Clipping from below
Clipping Effect Intensity Standard-deviation
Clipping Model Original Signal Clipped Signal
Clipping Model Original Signal Clipped Signal
Clipping Effect Original Signal Clipped Signal
Clipping Model For Image Noise Original Signal Clipped Signal
Clipping Correction Ideal EstimationEstimation under Clipping Direct Transformation Inverse Transformation
Direct Transformation Clipped from Ideal
Direct Transformation Clipped from Ideal
Clipping Correction Ideal EstimationEstimation under Clipping Direct Transformation Inverse Transformation
Ideal EstimationEstimation under Clipping Direct Transformation Inverse Transformation
Clipping from above and below Intensity Spatial Index 1 0 Clipping from above Clipping from below
Correction Results Intensity Standard-deviation
Algorithm Overview Wavelet Analysis Input Image Smooth Region Segmentation Smooth Region Segmentation Level Sets Segmentation Level Sets Segmentation Local Estimation Clipping Correction Clipping Correction Global Fitting Global Fitting
Experiments Original y Observation z degraded by Poissonian and Gaussian noise with parameters χ = 100 (a = 0.01) and b = 0.042
Results Intensity Standard-deviation Reduce the influence of fine textures and edges Standard-deviation Intensity
Test Image Intensity
Test Image Intensity
Test Image Intensity
Test Image Intensity Test Image
Denoising Clipped Signals Foi, A.,.Practical denoising of clipped or overexposed noisy images., Proc. 16th European Signal Process. Conf., EUSIPCO 2008, Lausanne, Switzerland, August Original Signal Clipped Estimated Signal Spatial Index
Denoising Clipped Signals Noisy Image
Denoising Clipped Signals (FujiÞlm FinePix S9600 Camera), Denoised Result
Denoising Clipped Signals Denoised and Debiased Result