Biomedical Engineering Reference
In-Depth Information
I-rectangle must have negative weights. One example of the image with points of
interest is presented in Fig.
4.14
.
A multipurpose image recognition system was developed, which contains a
feature extraction subsystem based on the permutation coding technique and one-
layer neural classifier. The main advantage of this system is effective recognition of
different types of images. The system was tested on handwritten digit recognition,
face recognition, and micro-object shape recognition problems (we will describe
this task in micromechanics applications). The large number of features used by this
system permits us to obtain a good (complete) description of image properties, and
for this reason good recognition results could be achieved. All the features are based
on random neural network structures, making them applicable to a wide range of
image types. The large number of elements of the associative layer permits us to
transform the initial image to the linearly separable space. In this space, the one-
layer neural classifier works effectively. So, the advantages of the one-layer
classifier (simple training rules and fast convergence) make the system good for
different applications.
It was shown in [
7
,
8
] that large-scale classifiers having more than 100,000
neurons in the associative layer demonstrate high performance in image recogni-
tion. The main drawback of a perceptron-like classifier of this type (for example,
LIRA [
49
]) is its sensitivity to object displacements in the image. In this work, we
proposed a new image recognition system, which is insensitive to small displace-
ments of objects. The best result of this system on the MNIST database is 44 errors
(Fig.
4.13
). We investigated the influence of the neural classifier size on the
classifier performance. The performance is nearly twice as good when the number
of neurons in the associative layer of the classifier increases from 32,000 to
512,000. The main drawback of this system is its sensitivity to object rotation and
scale change. For this reason, it is necessary to use image distortions during the
training and recognition processes. Image distortions increase training and recog-
nition time. In our experiments, the training time was 44 hours on a Pentium IV 1.9
GHz computer, and the recognition time was 35 minutes for 10,000 images. To
decrease training and recognition time, it is necessary to develop mask generation
methods, which permit a decrease of the sensitivity to rotation and scale change of
the object in the image. This work will be done in the future.
Fig. 4.14 The face with
points of interest
