Biomedical Engineering Reference
In-Depth Information
Table 4.2 Recognition error
number obtained on the
MNIST database
Rectangle size
h
*
w
7*7
9*9
11*11
13*13
Run
1
58
53
56
55
2
55
54
50
51
3
56
54
44
56
Average
56,3
53,7
50
54
Fig. 4.12 The whole set of unrecognized handwritten digits
Fig. 4.13 The face example
from the ORL database
absence of glasses. As a rule, the algorithms are tested on the ORL database in the
following manner: five images of one person are used for classifier training and the
other five are used for testing. There are two modes to make this partition.
In the first mode, they use the first five images for training and the last five
images for testing. In the second mode, they randomly select five images for
training, and the rest of the images are used for testing. The first mode was good
for comparing the classifiers until classifiers with almost a 100% recognition rate
