Zhongliang Fu
2) If G(i)< T,, then set G(i)=0. Character group is segmented into n sub-blocks in G(i)=0.
3) From left to right, check the width D of each sub-block. If D is approximate to the width of character,
then the sub-block is corresponding to a character. If D is approximate to the half width of character, then check
next sub-block. If the width of next sub-block is approximate to the width of character, then current sub-block is
corresponding to a character. If the width of next sub-block is half the width of character, then the current sub-
block is merge into a character with next sub-block.
The algorithm can process linked character and the character composed of multi-sub-blocks.
5. CHARACTER IDENTIFICATION
The method for character identification must possess the ability to adapt shift, rotation, scaling of character. The
ANN scheme implementing the invariance to shift, rotation and scaling have:
1) The invariant feature of extracted pattern is used as network input.
2) The transformed pattern that is shift-, rotation- and scale- invariant is used as network input.
3) Constructing a network model that is shift- and rotation- invariant.
These schemes have respective merit and shortcoming. But the third scheme can more incarnate the mechanism
of human brain identifying pattern and is easy to be realized in computer. Usually adopted network model is full
connective three-order NN. But the network model has some obvious limitation. a) It doesn't utilize the
structure information of pattern. b) It can only process regular distortion. The tolerance to errors that are brought
by the complex distortion of pattern and noise is low. Its rule realizing shift- and rotation- invariance is that if
pattern is shifted and rotated, the shape of triangle structured with the random three points in this pattern is
steady. In fact, when irregular distortion of pattern and noise exists, the rule isn't hold. The research in
physiology shows that human brain cell isn't full connected. ANN used for simulating human brain must
possess the characteristic of human brain.
By researching the identification method of handwritten character with BP network, Keiji Yamade proves that
locally connected BP network can resist a noise and the distortion of pattern.
In the paper, 3D neural network introduces 2D local connection and composes of a compound NN and BP
network for character identification.
A compound network composes of three sub-nets as Fig. 2. First sub-net is a 2D locally connected three-order
network. It realizes self-associations of pattern. Second sub-net is a classing network. It finishes classing
patterns.
The model and learning algorithm of two sub-nets is as follows.
| / e à . > v — —— |
t connective | fs 2 | Class
: three order | ‘ f** f d geural :
: neural | ? /^* 7 : .
| /e » * ^ e / network code
e 9» »» 9 — network | uw E me Le
nxn array nxn array
Fig.2 Acompound neural network architecture
5.1 2D Locally Connected Three-order Network
2D locally connective three-order network is substantively a monolayer feedback network. Neurone is ordered
308 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
Fm"