CONCORDIA UNIVERSITY
Department of Electrical & Computer Engineering
 
   
 

Dynamic Range and Stability of NR-IQA

Developed by Meisam Rakhshanfar and Maria A. Amer

  
 

Abstract

There are some applications where the quality measurement is used as a relative number comparison. In such applications, the relative values of measured quality are compared to detect a quality change. Thus, we consider two properties to evaluate the NR-IQA performance: dynamic range and stability. NR-IQA can better highlight the changes in the quality when it has a higher dynamic range. Let us assume Idist is a low-quality noisy or blurred image and Iopt is a high qiality optimal image or the ground-truth. NR-IQA with a high dynamic range gives results such that Q (Iopt) >> Q (Idist). This feature provides us the ability to detect a major quality difference by comparing measured quality values. We propose the ratio of the highest quality to a defined degraded image (noisy or blurred) to measure the dynamic range DR as in,
(1)
Higher values of DR show a better contrast between high and low quality inputs. NR-IQA can be designed to have a very high DR. This can be done for instance by suppressing lower values of Q(.) and magnifying the higher values. However, increasing the DR may have the downside of decreasing the stability and increasing the sensitivity. To measure stability, let us assume a random process (such as AWGN) creates different images It from a ground-truth with (almost) same quality (e.g., same distortion type with same PSNR). The NR-IQA also should give close results with a relatively small variation. If the number of random samples is large, the normalized standard deviation can be used to measure stability ST as in,
(2)
 
 

Numerical Results

For a relative quality comparison, dynamic range DR and stability ST can be considered. Higher dynamic range provides a more contrast between low and high quality images which is useful in detection of significant quality changes. A NR-IQA should also be stable by providing similar results when both nature and amount of degradation are similar. We used (1) and (2) to find the dynamic range and stability. To create low-quality images we degraded images using AWGN with standard deviation (σa = 10) and 5 × 5 Gaussian blur with sigma of 1 and measured the DR under noise and DR under blur. Table 1 compares the dynamic range of all methods. Our method provides high dynamic range in noisy and blurry conditions. We used (2) to test the stability. We generated 10 different noisy images (It, t ∈ {1:10}) with same PSNR using AWGN with σa = 10, and we calculated the QI using all methods. Table 2 compares the normalized standard deviation of QI and the average. The proposed method provides stable estimates having high DR as in Table 1.

1. The following experiments show average DR for 10 selected images from TID2008, Peppers, and Barbara using two types of distortions.

BRISQUE CPBD JNB LPC S3 BIQI MetricQ SDQI
DR
AWGN		 1.72 0.83 0.79 1.00 0.93 1.36 1.49 1.67
DR
G-Blur		 1.47 2.36 1.76 1.12 7.93 1.12 1.42 1.48
 
 

2. Normalized standard deviation (ST) of SDQI using 10 noisy samples with same PSNR (28.1dB).

BRISQUE CPBD JNB LPC S3 BIQI MetricQ SDQI
Barbara 2.50 0.16 0.47 0.23 0.38 0.34 0.28 0.41
Kodim05 0.58 0.15 0.82 0.14 0.35 0.36 0.41 0.33
Kodim06 1.16 0.20 0.89 0.23 0.27 0.33 0.33 0.50
Kodim07 1.02 0.19 0.79 0.14 0.48 0.29 0.38 0.23
Kodim10 1.23 0.18 0.83 0.18 0.22 0.20 0.70 0.61
Kodim17 1.43 0.19 1.21 0.19 0.41 0.21 0.64 0.42
Peppers 0.99 0.21 1.34 0.31 0.39 0.69 0.37 0.52
Average 1.27 0.18 0.91 0.20 0.36 0.35 0.44 0.43