Databases Reference
In-Depth Information
T A B L E 7 . 4
Prediction error entropies and
corresponding file sizes obtained using
the cap prediction.
Image
Residual Entropy
cap
Sena
3.41
27,934
Sensin
3.81
31,252
Earth
3.20
26,238
Omaha
6.03
49,937
before a prediction can be generated. If more than five matches are observed, the algorithm
takes the average of the pixels that have the matching contexts as the predicted value. If the
larger context cannot garner five or more matches, the algorithm shifts to the next smaller
context in a manner similar to
ppm
. If there are not sufficient matches for the smallest context
allowed, the algorithm uses the version of the median adaptive predictor used in JPEG-LS as a
default. Note that the decoder does not have to be informed about which context is being used
as the context selection is based on information available to both the encoder and decoder.
Using this approach, we obtain the results shown in Table
7.4
.
7.6 Multiresolution Approaches
Our final predictive image compression scheme is perhaps not as competitive as the other
schemes. However, it is an interesting algorithm because it approaches the problem from a
slightly different point of view.
Multiresolution models generate representations of an image with varying spatial resolu-
tion. This usually results in a pyramid-like representation of the image, with each layer of the
pyramid serving as a prediction model for the layer immediately below.
One of the more popular of these techniques is known as HINT (Hierarchical INTerpola-
tion) [
91
]. The specific steps involved in HINT are as follows. First, residuals corresponding
to the pixels labeled
in Figure
7.2
are obtained using linear prediction, and they are trans-
mitted. Then, the intermediate pixels
are estimated by linear interpolation, and the error
in estimation is then transmitted. Then, the pixels
X
are estimated from
(
◦
)
and
◦
, and the
estimation error is transmitted. Finally, the pixels labeled
are estimated from
known neighbors, and the errors are transmitted. The reconstruction process proceeds in a
similar manner.
One use of a multiresolution approach is in progressive image transmission. We describe
this application in the next section.
∗
and then
•
7.6.1 Progressive Image Transmission
The last few years have seen a very rapid increase in the amount of information stored as
images, especially remotely sensed images (such as images from weather and other satellites)
and medical images (such as computerized axial tomography (CAT) scans, magnetic resonance
