Image Processing Reference
In-Depth Information
Figure 4.2
Comparison of results of applying the filter design on purely additive noise with
optimal filter for the image shown.
the training set vary from the image to be filtered, then the results will be suboptimal.
In the extreme cases, there will be some strange effects.
A filter is generally only valid on the test set for which it has been trained. Con-
sider the test image shown in Fig. 4.1. It is corrupted by 10% noise with additive
and subtractive properties. Figure 4.2 shows the results of filtering this image using
the filter previously trained on the additive noise example in Chapter 3. A compari-
son is made with the results of filtering using the optimum filter for this image.
The optimum filter reduces the number of pixels in error from 6393 to 1462,
whereas the additive noise filter leaves 2436 pixels still in error. The nature of the
