Image Processing Reference
In-Depth Information
Figure 3.11
Results of resolution conversion. The upper image was originally scanned at
300 dpi. The lower image shows the results of filtering by a 17-point window trained on hun-
dreds of images to convert from 300 to 600 dpi. A hardware implementation takes less than
100 gates. Reproduced from Loce
2
where further details may be found.
original document was scanned at 300 dpi or if the resolution of the text font is low.
Filters of the type described above may be trained to carry out resolution conver-
sion from lower to higher resolution. An example of resolution conversion taken
from Loce is shown in Fig. 3.11.
2
A 17-point window was trained on hundreds of
examples of low- to high-resolution images. The example in the figure was imple-
mented in hardware with fewer than 100 gates. It can be seen that the upper image
scanned at 300 dpi has very jagged edges. After filtering, the resolution was con-
verted to 600 dpi and the edges are much smoother.
In order to test the robustness of the approach, a filter trained on a standard
western alphanumeric font was applied to kanji characters shown in Fig. 3.12. It
can be seen that the conversion works equally well even though these characters
had not been seen in training. This suggests that the image statistics for many types
of characters are very similar. In the above examples, there is no linear equivalent
filter to solve these problems.
3.3 Summary
This chapter has shown how the table of observations may be used for a number of
tasks related to filter error. In particular it has shown how the error after filtering
with the optimum filter may be calculated. It has also shown how the error between
filters may be compared. Other properties of the filter such as whether they are ad-
ditive or subtractive may also be deduced. Lastly, a number of practical examples
