18.4   Normalization Filter

The Normalization Filter connects the image to the underlying neural nets.

Let the underlying neural net be 100x100: if an image is larger than 100x100, say 350x230, then this image will be reduced to 100x100 or smaller.

When reducing images, a scaling factor can be introduced easily. Although scaling symmetry can compensate for this scaling factor, scaling symmetry is computationally expensive.

It is important to know that the Reduction Filter will match the selected underlying neural net, therefore, the behavior of the Reduction Filter not only depends on the selection of this filter itself, but also depends on the NeuralNet Filter chosen.

Figure 18.5  Selecting Reduction Filter.

There are several ways to reduce images:

•    Integer,
•    Real, or
•    All images are reduced by the same amount.

Integer Reduction

Images are reduced by an integer factor to maximally fit 100x100 without distortion. For example, a 350x230 image will be reduced to 87x57.

Real Reduction

Images are reduced by a real number to maximally fit 100x100 without distortion. For example, a 350x230 image will be reduced to 100x65.

Within each type of reduction, there are 3 more settings. Assume a 3x3 pixel array is reduced to 1 pixel,

•    Avg: Assign the average of the 3x3 pixel array to the new pixel;
•    Max: Assign the maximum of the 3x3 pixel array to the new pixel; or
•    Min: Assign the minimum of the 3x3 pixel array to the new pixel.

To select the Reduction Filter, use the fourth drop down list. The Reduction Filter has seven parameters.

Segment Cut

This parameter deals with the edges of the segments in the images. The Segment Cut parameter ranges from 0 to 12. The larger this parameter is, the smaller the segment the ImageFinder will use. The possible settings in the user interface are: 0, 1, 2, .., and 12.

Size Cut

In some applications, the users only want to search images of certain dimensions and ignore other images. An example is given below:

In this example, the two stamps belong to two different classes based on the image dimension alone.

The Size Cut parameter ranges from 0 to 9. If the setting is 0, this parameter will be ignored.

•    If the parameter is 1, then the longest edge of the image to be considered must be at least 100, but less than 199.
•    If the parameter is 2, then the longest edge of the image to be considered must be at least 200, but less than 299; … 

Border Cut

The Border Cut parameter ranges from 0 (no cut) to 9 (18% border cut). For some images (see the picture below), you might want to get rid of the sections of images close to the borders. To get rid of the border section, use the Border Cut.

The possible settings in the user interface are: 0, 1, 2, .., and 9.

Assume an image is (0,0; 1,1),

•    setting Border Cut to 1 means the ImageFinder will look at the section (0.02, 0.02; 0.98, 0.98);
•    setting Border Cut to 2 means the ImageFinder will look at the section (0.04, 0.04; 0.96, 0.96); … .  

Look-At Area

The Look-At Area is the area the ImageFinder will use in a matching operation. A 100 x 100 window specifies a whole image. If an integer Reduction Filter is used, the actual area can be less than 100x100.  

Four numbers specify the Look-At Area:

    (x, y, w, h)

(x, y) are the coordinates of the upper-left corner and (w, h) are the width and height of the Look-At window.

To use this Look-At window, enter (x, y, w, h) to the 4 text boxes.

Copyright (c) 2006 - 2007 Attrasoft. All rights reserved.

gina@attrasoft.com