|
Images formed by different systems are often noisy which makes filtering a typical operation of image pre-processing.
In many research papers, filter performance is analyzed for a limited number of standard test images and noise
variances. Here we use a recently created color image database TID2008 that allows assessing filter efficiency for 25
color images corrupted by noise with different values of variance, both i.i.d. and spatially correlated. Besides, this
image database serves the purpose of evaluating different quality metrics including those able to characterize visual
quality of original and processed images considerably better than conventional MSE and PSNR. The study is carried out
for filters based on discrete cosine transform (DCT) able to suppress both i.i.d. and spatially correlated noise depending
upon a way of threshold setting. It is shown that improvement of PSNR (IPSNR) due to filtering is very close for R, G,
and B components of color images and this improvement depends on image content. IPSNR reaches 9 dB for quite
simple images and it is only about 1 dB for highly textural images if initial PSNR=30 dB. Note that IPSNR is larger if
the original PSNR is smaller. The visual quality metric PSNR-HVS-M is studied as well. The metric PSNR-HVS-M
becomes larger due to filtering but in smaller degree than PSNR does. We demonstrate that it is possible to forecast
whether or not visual quality can be improved due to filtering or to detect in advance highly textural images for which
filtering can be not efficient enough. The provided output MSEs are also compared to potential limits calculated
according to the recently proposed methodology. It is demonstrated that for highly textural images the DCT filtering
with 8x8 full overlapping blocks and hard thresholding provides output MSE close to potential limits. The provided and
limit MSEs differ from each other by about 10%. For simpler images, the provided and limit MSEs can differ by
1.5...2.5 times. Analysis is also carried out for spatially correlated noise. It is shown that efficiency of filtering in this
case is lower.
|
