Information Technology Reference
In-Depth Information
neurons [31]. In this study, we employ a DCT based approach to model the con-
trast sensitivity function (CSF) and masking effect as in [23]. The quality value is
computed according to a modified PSNR by excluding the imperceptible distor-
tion because of the contrast sensitivity and masking effect from the computation of
PSNR. This method is called PHVS, and readers can refer to [23] for details.
4.2 Performance Evaluation of Image Quality Metrics in
D-Cinema
According to the sizes of tested images and the screen in our digital cinema setup,
as well as the human fovea acuity angle (2º), the block size in this study was set to
16 according to equation (4), and the threshold
T
was 10% calculated by equation
(5). A reference image and its distorted image (JPEG 2000 compressed) were
divided into different blocks with sizes 16×16. Subsequently, all blocks in the ref-
erence image were sorted in a descending order according to their standard devia-
tions, and 10% of all blocks with highest standard deviations were selected to
compute the quality of distorted image. A quality value in each candidate block
was calculated by PSNR, SSIM, or PHVS. The mean of the quality values over all
the candidate blocks was taken as an overall quality of the distorted image.
To evaluate the performance of the proposed approach, four evaluation criteria
were used. As aforementioned, a nonlinear mapping operation in equation (3) was
performed between the metric results and the subjective DMOS values. RMSE,
Pearson correlation coefficient, and Spearman rank order correlation coefficient
can be computed between the mapped metric results and the DMOS values. In ad-
dition to these three criteria, another criterion, outlier ratio relating to the predic-
tion consistency of a metric, can be obtained, because our subjective quality
experiment provided standard errors of the subjective quality results. The outlier
ratio is defined as the ratio of the number of outlier point images compared to the
total number of the tested images, in which an outlier point image is detected if it
satisfies the following condition:
DMOS
−
DMOS
> ⋅
2
SE
(6)
P
where
SE
denotes the standard error value.
In our experiments, the original methods of PSNR, SSIM, and PHVS were per-
formed with respect to the subjective image quality assessment in the digital cin-
ema setup, and four evaluation criteria were computed. To validate the proposed
approach, we used two methods as follows:
1) The first method was to divide the image into different 16×16 blocks, and these
three metrics were computed in all blocks. The mean over all blocks in the image
was taken as the quality metric for this image.
2) The image was still divided into different 16×16 blocks in the second method,
however the metrics were only performed for those blocks with high contrast lev-
els, as described above. The image quality was computed by the mean over these
