Summary: 862 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS--II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 48, NO. 9, SEPTEMBER 2001
ROBUSTNESS OF FILTERS
FILTERING RESULTS USING FILTERS DESIGNED BY IDEAL TRAINING
better visual performance than the other filters. We can also recognize
that the 3 2 3 window size is insufficient for such image processing.
C. Robustness of M-LWOS Filters
Both canonical LWOS and M-LWOS filters are designed by training.
In practical cases, the original images of corrupted images (i.e., input
images) are not available. Thus, training images are different from cor-
rupted images. It is necessary to assess the robustness of filters designed
by training. Both the canonical LWOS filter and the M-LWOS filter
(P = 9) are designed by using Lena which is corrupted by mixed noise
, p) = (400,8). We process six corrupted images as listed in
Table II by using these two filters. Table III shows the results for filters
which are designed ideally (using corrupted images and corresponding
The M-LWOS filter is superior to the canonical LWOS filter for all