CONCORDIA UNIVERSITY
Department of Electrical & Computer Engineering
Video Processing (VidPro) Group, Dr. M. Amer
 
 

Estimation of Gaussian, signal-dependent, and processed noise in Image and Video Signals

Meisam Rakhshanfar and Maria A. Amer
IEEE TIP: published September 2016
Contact: amer att ece.concordia.ca

  

Software (IVHC)

   Download the MATLAB code and related package here ✓ Download the Python package (Linux, Mac, Windows) here


 

Abstract

We propose a method to estimate real image and video noise including white Gaussian (signal-independent), mixed Poissonian-Gaussian (signal-dependent), or processed (non-white and frequency-dependent). Our method also estimates the noise level function (NLF) of these types. We do so by classification of intensity-variances of image patches in order to find homogeneous regions that best represent the noise. We assume the noise variance is a piecewise linear function of intensity in each intensity class. To find noise representative regions, noisy (signal-free) patches are first nominated in each intensity class. Next, clusters of connected patches are weighted where the weights are calculated based on the degree of similarity to the noise model. The highest ranked cluster defines the peak noise variance and other selected clusters are used to approximate the NLF. The more information, such as temporal data and camera settings, we incorporate, the more reliable the estimation becomes. To account for processed noise, (i.e., remaining after in-camera processing), we consider the ratio of low to high frequency energies. We address noise variations along video signals using a temporal stabilization of the estimated noise. Objective and subjective simulations demonstrate that the proposed method well outperforms, both in accuracy and speed, well-known noise estimation techniques.

 
 

Numerical results for synthetic noise (from the paper)

1. The following experiments show the absolute of estimated error in average for 14 different images, where AWGN with different variances is added to ground-truth images.

Sigma = 4 Sigma = 8 Sigma = 12 Sigma = 16
Absolute of error 0.22 0.15 0.14 0.15
 
 

The following experiments show the absolute of estimated error in average for 14 different images, where processed AWGN with different variances is added to ground-truth images. We used isotropic (Gaussian blur) and anisotropic (bilateral filter) process to make the noise spatially correlated (frequency-dependent).

Processed Noise σa = 10 σa = 15
Noise is processed by isotropic filter
Gaussian blur sigma σGB= 0.45 σGB= 0.5 σGB= 0.55 σGB= 0.45 σGB= 0.50 σGB= 0.55
Absolute of error 0.18 0.19 0.23 0.24 0.27 0.33
Noise is processed by anisotropic filter
Bilateral sigma σBL = 0.5σa σBL = σa σBL = 2σa σBL = 0.5σa σBL = σa σBL = 2σa
Absolute of error 0.24 0.28 0.38 0.27 0.32 0.44

Visual Results

The following visual comparison shows left the REAL-noisy and right the denoised using IVHC as the noise estimator.

Noisy
 Denoised

This work was supported jointly by TandemLaunch Inc., wrnch Inc., and Mitacs Canada. Some images are courtesy of wrnch Inc.
The authors are with the Electrical and Computer Engineering Department, Concordia University, Montreal, QC, Canada.
Contact: amer AT ece.concordia.ca
Copyright (c) 2014-2016 Concordia University and wrnch Inc. All Rights Reserved.