The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
Table 3: Distance for JPEG 2000 at 1 bpppb.
Degradation type
Deg. param.
Distance
# of misclass.
White noise
50
0.493597
112
White noise
100
0.492287
163
White noise
200
0.520257
255
White noise
1000
0.884674
634
Spectral smoothing
3
1.26847
262
Spectral smoothing
5
0.634936
166
Spectral smoothing
7
0.376314
123
Spatial smoothing
13
1.50541
4248
Spatial smoothing
15
1.22905
3778
Mixed smoothing
11
1.91304
4881
Gibbs
50
0.505195
698
Gibbs
100
0.510643
425
Table 5: Distance for JPEG 2000 at 0.5 bpppb on moffett4 image
with degradation on moffett3.
Table 4: Distance for a white noise of variance 100 on moffett4
image with degradation on moffett3.
Degradation type
Deg. param.
Distance
# of misclass.
White noise
50
0.107652
112
White noise
100
0.0435569
163
White noise
200
0.139344
255
White noise
1000
0.705412
634
Spectral smoothing
3
1.56918
262
Spectral smoothing
5
0.908723
166
Spectral smoothing
7
0.608943
123
Spatial smoothing
13
1.42293
4248
Spatial smoothing
15
1.12790
3778
Mixed smoothing
11
1.91772
4881
Gibbs
50
0.159746
698
Gibbs
100
0.198265
425
JPEG 2000
0.5
0.830507
450
JPEG 2000
1.0
0.449398
142
Degradation type
Deg. param.
Distance
# of misclass.
White noise
50
0.400136
112
White noise
100
0.419496
163
White noise
200
0.478324
255
White noise
1000
0.920144
634
Spectral smoothing
3
1.38425
262
Spectral smoothing
5
0.738476
166
Spectral smoothing
7
0.455491
123
Spatial smoothing
13
1.50678
4248
Spatial smoothing
15
1.22396
3778
Mixed smoothing
11
1.93958
4881
Gibbs
50
0.424221
698
Gibbs
100
0.407225
425
JPEG 2000
0.5
0.565506
450
JPEG 2000
1.0
0.129164
142
We also apply the method for a JPEG 2000 compression on mof-
fett4 at a bitrate of 0.5 bpppb. Distances are presented on ta
ble 5. The distance successfully identifies the degradation as be
ing JPEG 2000. The mojfett4 image is more uniform, thus easier
to compress. This property explains that degradations on mof-
fett4 for a bitrate of 0.5 bpppb are similar to those on mojfett3
at 1.0 bpppb. The number of misclassified pixels estimated by
this method is 142, the real value being 82. There is a lack of
reliability in this situation.
In the situation where the method is applied to a different image,
it is difficult to evaluate precisely the impact of the degradation
on the application. However, this method successfully identifies
the nature of the degradation caused to the image. It would be
worthwhile to detail these results on a greater number of images
and applications.
Interest compared to traditional SNR
SNR, as well as MSE or PSNR, which are derived measures, do
not reveal the nature of the degradation and do not allow inferring
the impact on applications. This fact is illustrated on figure 9. Di
agrams are plotted for different degradations leading to the same
SNR. The five degradations are applied to the moffett3 image to
provide a SNR equal to 30 dB. We can clearly see from the dia
gram that even if these degradations led to the same SNR, their
characteristics are completely different.
For the same image using only the SNR to measure the quality
(here for a SNR of 30 dB), we cannot make the difference be
tween
a white noise with a variance of 2000: 1054 misclassified
pixels;
a spectral smoothing with an attenuation parameter of 4: 207
misclassified pixels;