Biomedical Engineering Reference
In-Depth Information
Fig. 3.8 The scheme of 16
distortions
3.2.2.4 Recognition procedure
To examine the recognition rate, the test set of the MNIST database was used. This
test set contains 10,000 images. Coding and calculation of neuron activity were
made by the same rules as by training, but the value
T
E
(reserve of robustness) was 0.
The recognition process for the new classifier differs from the previous ones. In
this version, we use distortions in the recognition process, too. There is a difference
between implementation of distortions during the training session and during the
recognition session. During the training session, each new position of the initial
image produced by distortions is considered a new image, which is independent of
other image distortions. During the recognition session, it is necessary to introduce
a rule for decision making. All the recognition results of one image and its
distortions must be used for receiving one result, which gives the class name of
the image under recognition. We have developed two rules for decision making.
Rule 1. According to this rule all the excitations of the
R
-layer neurons are
summed for all the distortions.
X
X
d
n
E
i
¼
a
kj
w
ji
(3.14)
k
¼
1
j
¼
1
