Recognition of Textures, Object Shapes, and Handwritten Words - Neural Networks and Micromechanics - page 115

with different radii turned at a certain angle, the intersections of segments, and so
forth. Each feature could be present or absent at the particular point of the image.
For feature k , the value y k ( ij ) = 1 is introduced if the feature is present at the point
( i, j ) of the image, and y k ( ij ) = 0 is introduced if it is absent.
For the recognition of handwritten words, we used as features the segments of
straight lines with specific lengths and orientations. In total, we used ten features.
To extract these features, the lines of handwritten symbols are thinned to one pixel
and then thickened to three pixels. Thus, the image of the symbol acquires the
standard thickness of lines. The presence of a feature is determined by the presence
of the straight line with the assigned length and orientation in the word symbols.
The separately written word was read by a scanner with a resolution of 300-400
binary points per inch.
A brief description of the algorithm for handwritten word recognition for the
limited dictionary consists of the following. The image obtained with the scanner is
treated with the following operations:

1. thinning the lines on the image;
2. thickening the lines on the image;
3. extracting informative features;
4. coding the features;
5. inputting the obtained codes into the neural network for training or recognition.

To realize these operations, we used the algorithms of cellular logic, which were
developed and tested by the authors [ 16 ].
Let there be an image window with the size of 3 3 pixels. The pixel coordinates

in the window at point ( i , j ) are shown in Fig. 6.8 . Let us describe the algorithms of
cellular logic.

6.3.1.1 The algorithm of line thinning

The algorithm is realized according to the following formulas:

x ij ¼ð x ij & ð x ij 1 U x ijþ 1 ÞÞ U ð x ij & x i 1 j 1 & x i 1 j Þ U ð x ij & x iþ 1 j 1 & x iþ 1 j Þ ;

x ij ¼ð x ij & ð x ijþ 1 U x ij 1 ÞÞ U ð x ij & x i 1 jþ 1 & x i 1 j Þ U ð x ij & x iþ 1 jþ 1 & x iþ 1 j Þ ;

x ij ¼ð x ij & ð x i 1 j U x iþ 1 j ÞÞ U ð x ij & x i 1 j 1 & x ij 1 Þ U ð x ij & x i 1 jþ 1 & x ijþ 1 Þ ;

x ij ¼ð x ij & ð x iþ 1 j U x i 1 j ÞÞ U ð x ij & x iþ 1 j 1 & x ij 1 Þ U ð x ij & x iþ 1 jþ 1 & x ijþ 1 Þ ;

(6.5)

( 1

if the pixel with coordinates ij belongs to the line

; ;

where x ij ¼

0 if the pixel with coordinates ij does not belong to the line.

;

x i - 1j - 1
x ij - 1

x i - 1j x i - 1j + 1

x ij x ij + 1

Fig. 6.8 Image window with
the size of 3

x i + 1j - 1 x i + 1j x i + 1j + 1

3 pixels