7.5.4.4   Basic Features

This dialog enables you to choose several basic features that affect the classification.

Pattern Width and Pattern Height

With the parameters Pattern Width and Pattern Height you set the fixed pattern size which is used for classification, that is, the size to which characters are scaled. This pattern size influences the parameter Gray Values as this, if activated, uses the interpolated gray values. Depending on how much the characters are scaled, or if they are not scaled at all, that is, how much Pattern Width and Pattern Height differ from the size of the character in the image, a different Interpolation method might be suitable for the application.

Setting a bigger size generally helps to distinguish more characters. If the value is, however, chosen to big, overspecification may be a problem. In this case the amount of time necessary for the training and the time necessary for the recognition will also increase.

Interpolation

The parameter Interpolation lets you choose the interpolation mode, that is the adaptation of characters in the image to the pattern size. It also influences the parameter Gray Values as this, if activated, uses the interpolated pattern. The most suitable interpolation method largely depends on the values that were chosen as Pattern Width and Pattern Height, as the scale factors.

Parameter	Effect	Usage
Constant (Default)	Bilinear interpolation in an image with a mean filter.	Recommended choice if characters are scaled down, but not by a large amount. This interpolation method achieves a high precision without requiring to much processing time.
Weighted	Bilinear interpolation in an image with a Gaussian filter.	Recommended choice if characters are scaled down by a large amount and a very high precision is required. Note that this interpolation method requires more processing time.
Bilinear	Interpolation, using the values of the 4 closest pixels in diagonal direction.	Can be used if characters are not scaled very much.
Nearest Neighbor	No interpolation is performed.	Fast but not very precise interpolation. Should only be used if the image is blurred.

For more information on this parameter, see also affine_trans_image.

Pattern Features Used for Classification

In this part of the dialog, three characteristics defining the classification of pattern features can be activated:

Activate

• Gray Values to use the gray values of the interpolated pattern,
• Symbol Region to use the symbol regions of the interpolated pattern, and
• Gradient to use gradient features instead of pixel patterns (8 directions sampled from a fixed 5x5 grid).

Symbol Features Used for Classification

In this part of the dialog, three characteristics defining the classification of symbol features can be activated:

Activate

• Ratio if the ratio (width to height) of a symbol should be used,
• Anisometry if the anisometry of the equivalent ellipse of the symbol shape should be used, and
• Convexity if the convexity of the symbol shape should be used.