| Summary: | 碩士 === 國立成功大學 === 電機工程研究所 === 82 === In this paper, a system included image preprocessing and neural
networks is proposed. The various function units of the image
processing are used to obtain an invariant image representation
in the beginning of the system. The space of the neural
networks weights can be reduced by using the reduction of
the feature dimension before the preprocessed feature applied
to the networks.Then, several kinds of the neural models are
proposed for pattern recognition : (1) distributed associative
memory (DAM), (2) backpropagation network(BPN), (3)DAM combined
with BPN, and(4)BPN with the associative memory as initial
weights. In the case of (3), this hierarchical networks consist
of two levels of neural networks. In the low level, a DAM
receives the output vectors of image preprocessing functions to
create a system which recognizes pattern regardless of changes
in scale or rotation. The higher level is a two- layers BPN
which recives the recalled information from the memorized
database of the lower level. This neural networks use a BPN
after the DAM can raise the recognition ratio in comparison
with a DAM, and be faster than a BPN.In the case of (4), the
training of the BPN speeds up much because this neural networks
use a associative memory of a DAM as initial weights of the
first layer of te BPN. Experiment results show that the system
can recognize all the patterns correctly when the percentage of
the white noises is under under 20% for the case (3) and (4).
|
|---|
