Biomedical Engineering Reference
In-Depth Information
Table 4.1. Comparative results on the ORL database
Method
Error
Ref.
Year
PDNN
4.0%
[
26
]
1997
SVM + PCA coef.
3.0%
[
27
]
2001
Continuous n-tuple classifier
2.7%
[
28
]
1997
Ergodic HMM + DCT coef.
0.5%
[
29
]
1998
Classifier with permutation coding
0.1%
[
36
]
2004
Pseudo 2D HMM + DCT coef.
0%
[
30
]
1999
Wavelet + HMM
0%
[
25
]
2003
could ensure high-speed and low-cost image recognition. Therefore, it is necessary
to search for more general methods that would give sufficiently good results in
different recognition problems. There are general-purpose recognition systems, for
example, LeNet [
1
,
2
], Neocognitron [
31
-
34
], and Receptive Fields [
35
], but the
recognition quality of such systems is lower than that of special systems. It is
necessary to develop a general-purpose recognition system having quality compa-
rable to that of specialized systems.
Some years ago, we started to investigate a general purpose image recognition
system. In this system, we use the well-known one-layer perceptron as a classifier.
At the center of our investigations is the creation of a general-purpose feature
extractor, which is based on the concept of random local descriptors (RLDs). We
consider the terms “feature” and “descriptor” as synonyms. We intend to develop
the method of RLD creation, which could be used successfully for different types of
images (handwriting, faces, vision-based automation, etc.)
4.2 Random Local Descriptors
RLDs are based on two known ideas. The idea of random descriptors of images was
proposed by Frank Rosenblatt. In his three-layered perceptron [
37
], each neuron of
the associative layer plays the role of random descriptor of the image. Such a
neuron is connected to points randomly selected on the retina (input image) and
calculates a function from the brightness of these points. It is important to note that
the connections of the neuron are not modifiable during training. Another idea of
local descriptors is drawn from the discovery of Hubel and Wiesel [
38
,
39
], who
have proved that, in the visual cortex of animals, there are local descriptors that
correspond to local contour element orientation, movement, and so on. The discov-
ered local descriptor set is probably incomplete because, in the experiments of
Hubel and Wiesel, only those descriptors or features were detected that were
initially prepared to present to the animals. Probably, not all the descriptors
(features) that can be extracted by the visual cortex of the animals were investi-
gated. Rosenblatt's random descriptors could overcome this drawback, but the
application of these descriptors to full-size images decreases the effectiveness of
the application.
