The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
316
• Relative RMSE
RRMSE =
(3)
Note that, as the RRMSE enhances the importance of er
rors on small values, it is important to consider the partic
ular case of values below the sensor noise. Small values of
I(x, y, A) can be considered as random and are ignored for
the computation of the RRMSE.
• Spectral Fidelity [Eskicioglu and Fisher, 1995]
F\ = min |F . (4)
with
F{U,V) = 1 —
CliU-V)
Cl{U) •
Q(x,y) [Wang and Bovik, 2002]
Q(x, y ) = min {Q (/(•, -, A),/(-, -, A)) |.
Q(U,V)=;
( a U + cr v)(h L lJ F P’v)
where auv is the covariance between U and V.
(5)
(6)
(7)
All these criteria measure a distance between an original image
and a degraded version of this image.
Representing a combination of five values is a challenge and work
ing on a five-dimensional plot would not enable efficient assess
ment. A good way to represent these values is to use a star dia
gram (Fig 2) which gives a more intuitive vision than a classical
x-y representation in this case. The five axes of the diagram cor
respond to the five quality criteria. Scale for all the figures in this
paper are the same. For MAD, MAE and RRMSE, origin cor
responds to 0 (no degradation). The extremity of the axes corre
sponds to value 5000 for MAD, 40 for MAE and 0.1 for RRMSE.
For F\ and Q( x , y ), origin corresponds to 1 (no degradation), ex
tremity being 0.9 for F\ and 0.6 for Q( XtV ). These values were
found to provide a good differentiation between different degra
dations. These specific values are important for visualization and
comparison, they are not important by themselves, it is just neces
sary to use the same scales on the different figures. The shape of
the diagram is characteristic of the degradation as seen in figures
2, 3 and 4.
Parameters for the degradation are presented in [Christophe et al.,
2005]. Basically, this is the variance for the white noise, the filter
size for the smoothing, the scale factor for the Gibbs filter, and
the compression rate for JPEG 2000.
2.2 Shape characterizes the degradation
This representation is robust relatively to the amplitude of the
degradation. The shape is similar for a given degradation; the
degradation pattern is inflated when the degradation level increases
(Figs. 2-4). For example in Fig. 2, the innermost shape (green)
corresponds to a white noise with a low variance. When the noise
variance increases, the quality decreases and the quality diagram
dilates.
MAE
Figure 2: Quality for different values of additive white noise on
moffett4 image.
MAE
Figure 3: Quality for different values of spatial smoothing on
mojfett4 image.
MAE
Figure 4: Quality for different values of spectral smoothing on
moffett4 image.