Image Processing Reference
In-Depth Information
Chapter 4
How Do You Train the Filter for a
Task?
At this stage, the reader might be asking the obvious question of why do we need to
restore an image if the ideal original is available? In practice, a filter is used that has
been designed on a representative training set. This means that examples similar to
the image to be restored must be produced in some way. In the case of a fax ma-
chine, this is easy—a test image would simply be passed through the same process.
The same is true for resolution changing and OCR examples discussed in the previ-
ous chapter. For old film restoration and other processes, it can be more difficult to
recreate an ideal image for training. A section at the end of this chapter deals with
this subject in more detail.
The extent to which a filter trained on one image may be applied to another is
known as its
robustness
. If the statistics of either the noise or the image content of
Figure 4.1
Test image with 10% additive and subtractive noise (6393 pixel error).
29
