Biomedical Engineering Reference
In-Depth Information
images are centered and transformed to the gray scale. Many recognition systems
were tested on this database, and the results are available in the literature [
1
-
7
].
Some systems showed good results. The best result in [
1
] is 70 errors from 10,000
samples using the convolution neural network LeNet-4. In 2001, Belongie [
5
]
reported the result of 63 errors with a system based on the shape matching method
[
6
]. In 2001, we developed the neural classifier LIRA that showed 79 errors [
7
]. In
2002, we improved the LIRA [
8
] and repeated Belongie's result of 63 errors
(Part II). Further improvement of LIRA classifier permits us to obtain 55 errors
[
24
]. The new recognition system that we proposed gave the result of 44 errors.
Cheng-Lin Liu et al. in 2002 investigated handwritten digit recognition by combining
the eight methods of feature extraction with seven classifiers. They obtained 56
results for each tested database, including the MNIST database. The best result
obtained on the MNIST database was 42 errors [
9
]. This new method of image
recognition based on Permutation Coding Neural Classifier (PCNC) shows the
result of 49 errors, 0.49% as a mean value of three runs. The best result of this
method is 44 errors.
Automatic face recognition could be used in different security systems (for
buildings, banks, etc.), as well as for passport and other document verification.
Different approaches are investigated and proposed for solving this task [
10
-
20
].
They were tested on the database ORL (Olivetti Research Laboratory) [
19
,
21
-
23
].
We have tested our classifier on this database and obtained one of the best results.
The error rate was 0.1%.
The problem of micro-object shape recognition appears in microequipment
technology [
46
], where we use adaptive algorithms to improve the control systems
of micromachine tools. We will describe the results of microscrew shape recogni-
tion in the last part of the topic. For all the listed problems, we apply the same
PCNC algorithm.
At present, very good results are obtained using special recognition systems. Let
us, for example, compare the best results obtained in handwritten digit recognition
on the MNIST database and face recognition on the ORL database. The MNIST
database contains 60,000 handwritten digits in the training set and 10,000 hand-
written digits in the test set. There are different classifiers that have been applied in
the task of handwriting recognition [
1
,
2
,
5
,
6
,
8
,
9
,
24
]. The results are presented in
Table 3.1.
The ORL database contains 400 photos of 40 persons (ten photos of each person)
that differ in illumination, facial expression, and position. Five photos of each
person are used for training and the other five photos are used to test the recognition
system. The best results obtained on the ORL database are given in Table
4.1
from
[
25
]. As can be seen, almost all of the best recognition systems for the ORL
database differ from the recognition systems for the MNIST database. Some of
the systems use the same types of classifiers, for example, SVM and multilayer
neural networks, but the features extracted from the images in these cases are
different.
The great variety of recognition systems takes a huge amount of human work for
software development and complicates the development of special hardware that
