CONCORDIA UNIVERSITY
Department of Electrical & Computer Engineering

Video Processing (VidPro) Group, Dr. M. Amer


 
   
 

Sparsity Based No-Reference Image Quality Assessment for Automatic Denoising

Meisam Rakhshanfar and Maria A. Amer
Springer Journal Signal, Image and Video Processing: November 2017
Contact: amer att ece.concordia.ca

 
 

Abstract

In image and video denoising, a quantitative measure of genuine image content, noise, and blur is required to facilitate quality assessment, when the ground truth is not available. In this paper, we present a no-reference image quality assessment for denoising applications, which examines local image structure using orientation dominancy and patch sparsity. We propose a fast method to find the dominant orientation of image patches, which is used to decompose them into singular values. Combining singular values with the sparsity of the patch in the transform domain, we measure the possible image content and noise of the patches and of the whole image. We show that the proposed method is useful to select parameters of denoising algorithms automatically in different noise scenarios such as white Gaussian and processed noise. Our objective and subjective results confirm the correspondence between the measured quality and the ground truth. We show that the proposed method rivals related state-of-the-art no-reference quality assessment approaches.

Software

To get the MATLAB code Download here

Dataset

To get the Dataset Download here

 
 

Supplementary results



Numerical Results (from the paper)

1. The following experiments show QI of real noisy and denoised images using SDQI for sample images. SDQI shows quality improvement as it confirmed subjectively.

Room Painting1 Office Painting2 Tears Church Turbine Baby
Noisy 13.30 21.03 12.91 20.05 12.52 10.87 16.44 22.72
Denoised 15.20 27.58 15.78 26.98 14.71 13.87 18.02 26.91
 
 

2. Average of different metrics for denoised images from TID2013 using SDQI as a denoising parameter selector under different noise types. Ground-truth is PSNR and MSSIM.

SROCC
(PSNR)
KROCC
(PSNR)
SROCC
(MSSIM)
KROCC
(MSSIM)
PSNR MSSIM
AWGN 0.57 0.50 0.61 0.53 32.46 0.87
Spatially
Correlated
AWGN
0.63 0.57 0.70 0.65 31.23 0.85
Lossy
Compressed
AWGN
0.65 0.57 0.64 0.56 31.92 0.86

Visual Results

Each of the following visual comparisons show left the REAL-noisy and right the denoised part of sample images. We expect higher QI in the denoised output. Table 1 above confirms that.

Church
Office
Painting 1
Painting
Room
Tears
Turbine
Baby

General Distortions

3. SROCC and KROCC values for SDQI using the actual distortions in the TID2013 database and using the MOS as the ground-truth.

AWGN Gaussian
Blur
JPEG
Compression
JPEG2000
Compression
Spatially
Correlated
Noise
Impulse
Noise
Compression
+ Noise
Denoise
SROCC
(MOS)
0.86 0.90 0.53 0.76 0.69 0.86 0.84 0.76
KROCC
(MOS)
0.83 0.85 0.46 0.73 0.67 0.84 0.81 0.70
 
 

 
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@ece.concordia.ca and m_rakhsh@encs.concordia.ca
Copyright (c) 2014-2017 Concordia University and wrnch Inc. All Rights Reserved.