Binarization of Matrix Codes - Hierarchical Neural Networks for Image Interpretation - page 169

(a)
(b)
Fig. 8.20. Low-contrast degraded images (originals in Fig. 8.3(b)): (a) with random degrada-
tions; (b) output of the improved adaptive thresholding for the degraded images.
So far, only results for the high-contrast address window variant of the Data Ma-
trix codes have been reported. In the remainder of this section, results of binarizing
the low-contrast red ink code variant are described.
The 694 images were partitioned randomly into a training set of size 467 and a
test set of size 227. Since these images differ in brightness, contrast, and cell size
from the address window code variant, a retraining of the network was necessary.
Again, for each image a degraded version was produced. This time the images
were degraded only moderately since the image quality was already low. Only up
to nine vertical lines were added, and the amplitude of the background level as well
as the amplitude of the pixel noise was reduced by a factor of two. Figure 8.20(a)
shows degraded versions of the images from Fig. 8.3(b).
The desired outputs were produced using an improved adaptive thresholding
method that smoothed the image horizontally by applying a 1 / 4 (1 2 1) binomial
kernel prior to contrast stretching. This reduced the effects of the vertical dark and
bright lines present in the images. Figure 8.20(b) shows the binarized output of this
method for the degraded images.
The network was trained to iteratively reproduce the desired outputs not only
for the original images, but for their degraded versions as well. After training, the
network's behavior was very similar to the behavior of the network trained to bina-
rize the high-contrast address window variant of Data Matrix codes. The network
develops a stable representation of the cell structure which is used for binarization.
Figure 8.21 shows the average squared output changes and the average squared
difference to the desired output over time. The network converges quickly to an
attractor and stays there even when iterated further than the ten iterations it was
trained for. This attractor is close to the desired output. Generalization is good since
the curves for the training set are virtually identical to the test set curves.
The binarized outputs were evaluated by Siemens ElectroCom Postautomation
GmbH. For this experiment the recognition engine was queried a second time with
different parameter settings when an example was rejected in the first run. Table 8.1
summarizes the recognition performance for the entire dataset. It can be seen that