Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 140

A foreground pixel is removed when this counter is small, compared to its inten-
sity. Isolated pixels are always removed. Pixels with a counter of one are removed
only when their intensity is greater than 64. Pixels with a counter of two are removed
only when their intensity is greater than 128. Similarly, to weaken the inner part of
blobs, the intensity of a pixel is increased half the way towards 255 if the counter is
large, compared to its intensity.
Figure 7.5(c) shows the resulting images. While the removal of isolated pixels is
quite obvious, the weakening of blobs is most visible in the lower parts of the digits
four and in the three horizontal lines at the start of the last two blocks.
7.3.2 Normalization
The automatically determined regions of interest containing digits of the meter value
have a variable size, while the block classifier is a neural network with a fixed input
size. Hence, a mapping must be computed from filtered regions to the input image.
This mapping normalizes digit position and slant in order to simplify recognition. If
the normalization did not occur, the network would also have to learn translated and
slanted variants of the meter values. The size of the digit block is not normalized
because the image resolution is so low that an arbitrary scaling would produce sig-
nificant blur. In addition, it would require a reliable segmentation of the digits from
other objects, e.g. from the curved line delimiting the meter mark that is sometimes
included in the region of interest.
Slant Normalization. Due to misalignments of letters relative to the stamp dur-
ing metering, and relative to the camera during capture, some meter values appear
slanted in the rectangular region of interest. The slant normalization step estimates
this slant and corrects for it.
To estimate the slant, the center of mass of the dark foreground is computed in
the first step. It divides the image into a left and a right part. The centers of mass of
these two parts are computed next. They define a line that corresponds to the slant
estimate. Figure 7.6(a) illustrates this for three examples.
To correct for the slant, a vertical sheer transformation is used that makes the line
approximately horizontal. The sheering keeps the position of the center constant and
shifts columns only by an integer number of pixels to avoid blurring. Figure 7.6(b)
shows the resulting deslanted images for the three examples. Note that this normal-
ization has no effect if the estimated slant is relatively small.
(a)
(b)
Fig. 7.6. Slant normalization: (a) contrast enhanced image of slanted examples with markers
at the center of gravity, as well as at the left and the right center; (b) result of vertical sheering
that removes the slant.