Information Technology Reference
In-Depth Information
where M, N represent the number of rows and columns in the binary image. The
algorithm scans the output image (i.e. the image obtained after CA processing)
skipping all “white” pixels. If a pixel is “black” the algorithm is looking for its
lower and righter neighbors (two pixels) and if both are black then it is a new
upper left corner, the index ix of detected characters is incremented and the posi-
tion (Px, Py) and the size (Sx, Sy) of that character box are determined. They can
be stored in a list for further processing or they can be used immediately. In this
later case the function “FEATURES”, operating on the input image (i.e. the visual
field as it was extracted at the image sensor level) is called, to compute the feature
vector F to be submitted to the classifier. The classifier (function “RECOGNIZE”)
is then called to determine the content of the box.
There are many possibilities to classify the array of pixels within a box, most of
them based on neural networks. But since the boxes may have different sizes a
robust feature extractor should be employed, being capable to capture the useful
information in a fixed, predefined number of features. The basic idea is exempli-
fied in Fig. 8.5.
The box containing a character is divided into six rectangular blocks and each
block is assigned a numerical label associated with a feature component
f
,
K .
1
6
Fig. 8.5. The schematic of a robust feature extractor
Search WWH ::




Custom Search