Next: References Up: Title
Page Previous: Methods Index:
Full Text Index Contents: Conference
Page
Results and Conclusions
In order to test the performance of the proposed compression
procedure we are using a subjective criterion of visual appearance and
an objective criterion, the rate/distortion factor. We are comparing
the wavelet compression with JPEG compression in term of those two
criteria. In figure 8
a detail of a_vm1480 slice (male thorax) is showed with the corresponding
details of compressed images at different compression rates. One can easily
view the visual distortion is not perceptible till compression rates of
100:1 and the space saving is very attractive. We have measured the
image distortion as the root mean squared error between the compressed
image and the original one for each RGB color channel. As can be seen for
EZ wavelet approach, the residual distortion is lower than JPEG one
at each compression rate.
|
Compression Ratio
|
Size (KByte)
|
RMS(R,G,B) Wavelet
|
RMS(R,G,B) JPEG
|
|
10:1
|
~500
|
(0.0071,0.0069,0.0077)
|
(0.0106,0.0070,0.0109)
|
|
40:1
|
~190
|
(0.0129,0.0125,0.0123)
|
(0.0155,0.0129,0.0156)
|
|
85:1
|
~87
|
(0.0181,0.0186,0.0172)
|
(0.0213,0.0191,0.0218)
|
|
200:1
|
~39
|
(0.0246,0.0265,0.0242)
|
(0.0344,0.0311,0.0362)
|
|
300:1
|
~24
|
(0.0290,0.0317,0.0282)
|
(0.0410,0.0427,0.0536)
|
|
450:1
|
~17
|
(0.0323,0.0359,0.0315)
|
(0.0714,0.0636,0.0837)
|
|
600:1
|
~12
|
(0.0359,0.0392,0.0336)
|
(0.0921,0.0848,0.1023)
|
Table I. Comparison in between EZ wavelet and JPEG compression
in term of rate/distortion.
In figure 9 the
above results are showed in term of image and channel difference between
original and compressed image. One can note the residual error of JPEG
is higher than one obtained with EZ wavelet. Relevant residual structures
are visible in both difference images which are related to the edges of
structures and inside the spine section where a texture with high-frequecy
content is present. In figure 10
the image difference are plotted in 3D for each color channel. In the left
side results for compression ratio of 40:1 are shown whereas in the right
side the compression ratio is 200:1. One can expect that the structure
dimension which can be recover with reliability is directly related to
filter used to image transform anf to compression rate. Currently we are
involved in the problem of quantifing the size of this dimension in order
to choice the compression rate and filters according to anatomical districts
being saved into lossy compression. This approach could customize the image
compression according to different specialties which require to pay attention
to the structures they are interested in and disregarding other anatomical
features.
We have also shown that progressive coding are easily
generated through Embedded Zerotree algorithm. The user is allowed to browse
only a low-resolution approximation which is sufficient for deciding if
the image needs to be decoded or received in original full size. Future
developments are related to the evaluation of the performance of the wavelet
filters when analyzing VHD color image (Daubechies, Antonini, ../orthogonal
versus biorthogonal filters). Another issue will be addressed which involve
the software architecture to utilize. Browsing and visualizing at the same
time would require applet or plug-in Java code to send to the user station.
Yet due to the fact that Java is interpreted this approach to image synthesis
is, for the time being, very slow. Unlike Java, a different solution is
to provide the user with a compiled (Windows/X) application and the image
synthesis, very rapid, has to be done off-line with respect the browsing.
Next: References Up: Title
Page Previous: Methods Index:
Full Text Index Contents: Conference
Page
