Image Processing Reference
In-Depth Information
Quality of an image is an intangible subject for the common observer. Given
two images, similar in content but significantly different in quality, one can easily
recognize a relative quality. When the images are meant to be viewed by a human
observer, the best method of quantifying the visual quality is through subjective eval-
uation [173, 184]. However, these quality measures depend on psychovisual factors,
and therefore, such a subjective evaluation is difficult to reproduce and verify [136].
Furthermore, it turns out to be a too inconvenient, expensive, and time-consuming
option. Therefore, it makes sense to develop a certain set of measures to assess the
visual quality of an image objectively. The objective performance measures appear
as a valuable complementary method to the subjective evaluation [136]. Wang et al.
have enlisted three applications of designing an objective quality metric [184, 186]:
•
to monitor image quality for control systems,
•
to benchmark and compare image and video processing algorithms, and
•
to optimize algorithms and parameter settings for image processing systems.
The image quality measures are broadly divided into two categories—(
i
) full-
reference measures, and (
ii
) no-reference measures. In the former category, an orig-
inal (noise- and distortion-free) image known as the reference image or the ground
truth is available. The quality of an image (result of fusion in this context) is mea-
sured as its deviation from the original or the ideal image. An overview of such
measures for assessing the fidelity of grayscale and multispectral images can be
found in [9]. However, these measures require an ideal reference image. Practically
it is not possible to have the
ideal
fused image. (Had it been available, fusion is no
more needed!). Some performance measures quantify how closely the fused image
is related to the constituent input images. Li et al. have composited an ideal-like
image by manual cut-and-paste procedure from a set of multi-focus images [104].
However, this method is tedious, and is not applicable in general. Once the ideal,
or the reference image has been defined, a metric such as root mean square error
(RMSE) can be employed to compute the deviation, and thus, the quality as given
by Eq. (
2.4
).
1
2
1
/
2
XY
x
RMSE
=
(
I
ref
(
x
,
y
)
−
F
(
x
,
y
))
,
(2.4)
y
where
I
ref
is the reference image.
X
and
Y
indicate the image dimensions. The mea-
sure of MSE (or RMSE) is also related to the peak signal to noise ratio (PSNR).
These measures are easy to calculate, and are mathematically convenient to inte-
grate into further processing. Although the MSE-based measures provide an elegant
physical interpretation, these are not well correlated with the perceived quality of the
image [185].
Xydeas and Petrovic have proposed an objective measure to assess the quality
of fusion based on the edge information [194]. This measure is quantified on the
basis of
edge
information present in each pixel as it constitutes an important visual
information for the human visual system. Their measure uses the Sobel edge operator