Information Technology Reference
In-Depth Information
degraded
original
degraded
original
1
1
0.75
0.75
0.5
0.5
0.25
0.25
0
0
0.94
0.95
0.96
0.97
0.98
0.99
1
0
500
1000
1500
2000
(a)
(b)
confidence
squared output error
Fig. 8.17. Data Matrix test set recognition. (a) confidence versus recognition error; (b)
squared output distance versus recognition error.
In Figure 8.16 the confidence of all training and test examples at iteration 10 is
shown versus the squared distance to the desired output. Here, a clear distinction
between the original images and the degraded images becomes obvious. For the
original undegraded images, outputs with higher confidence and lower distance are
produced by the network than for the degraded images. In fact, none of the originals
has a lower confidence than any of the degraded images.
Again, the plots for the training set and the test set as well as their average
confidences and output errors, summarized in Fig. 8.16(c), are very similar. One
can also observe that the confidence is anti-correlated to the output error and can
hence be used to reject ambiguous examples that could lead to recognition errors.
For all examples, the output of adaptive thresholding as well as the output of
iterative binarization has been presented to a Data Matrix recognition engine. This
evaluation was done by Siemens ElectroCom Postautomation GmbH. It produces
for each recognized example the percentage of error correction used. This value
will be referred to as recognition error. It is set to one if an example could not be
recognized at all.
Both methods recognized 514 (99.8%) of the 515 original examples. The de-
graded images were harder to recognize. From the adaptive thresholding output 476
(92.4%) images could be recognized, while the system recognized 482 (93.5%) ex-
amples when the iterative binarization method was used. For comparison, the system
could only recognize 345 (70.0%) of the 515 degraded examples if a simple global
thresholding method was used for binarization.
Figure 8.17 shows for the iterative binarization method, the recognition error of
all test examples plotted against the confidences and against the squared distances to
the desired output. In both cases, the relation between the two quantities is not obvi-
ous. The degraded images that could be recognized do not need a higher percentage
of error correction than the original images. However, all examples that could not be
recognized are degraded images having a relatively low confidence and a relatively
high distance.
In Figure 8.18(a) the recognition error of the adaptive thresholding and the iter-
ative binarization method are compared for the test set. The performance of the two
Search WWH ::




Custom Search