Zhongliang Fu 
  
Character segmentation is the base of character identification. Because a character of number code is not 
unattached and connective sometimes, this brings character segmentation difficulties. Although the process 
speed of conventional region-growth method is fast, but they only segment unattached and connective 
characters. Oscillator neural network (OSNN) goes on visual sense model, but it also only process unattached 
and connective objects, and the run-time is long. 
At present, character identification mostly adopts template matching method and feature statistics classing 
method. Template matching method can’t adapt some distortion of characters (for example rotation, scale, local 
distortion, etc) and is badly affected by noise. Its computing quantity is great. For the feature statistic classing 
method, choosing feature is difficult. Classing result is affected by statistic distributing rule and noise. Structure 
information of characters can’t be utilized. 
. Aiming at above problems, the paper first discusses dynamic threshold method based on statistic and structural 
feature of gray level, and proposes a local contrast method taking into account stroke width feature for fast 
extraction of character. Then, characters segmentation is realized with an improved projection method. In the 
end, artificial neural network (ANN) method for character identification is discussed. A new method based on a 
compound neural network for character identification is proposed. 
2. CHARACTER EXTRACTION 
The purpose of character extraction is to separate the pixels in character stroke and other pixels. 
Suppose the width of character stroke is W, gray value of pixel (x,y) is f(x,y), the checked pixel is Q, Point P; 
(i=0,1, ..,7) is in neighborhood of point Q. They distributes as Fig 1. The average gray value in neighborhood of 
P; is A;. The size of neighborhood is QW+1) x (2W+1). 
— Pe 
PPE Tes MP Nes ^ 
Po |] EA n1] D / \ DI fn 
COND I [v d [ 
x : \ 0 ) X 
PS | ” 
Le of S P7NJ ps ES 
PS ~~ E 
(a) (b) (c) 
Fig. 1 neighborhood distribution 
Suppose N=(2W+1) x (2W+1), then: 
A= S So N (1) 
i--wj--w 
If the gray value of the pixel in character stroke is great than one of other pixel, then 
lif f(x, y)—A, ^ T,; 
0, otherwise 
LP) 4 (2) 
Where, T, is a threshold. f(x, y) may be replaced with the average gray value in neighborhood of point Q. 
If the pixel Q is in character stroke can be confirmed by following formula. 
Lif y [UP)AUP,)]=k 
B( x y) = i=0,1,2,3 ( 3) 
0, otherwise 
Where B(x,y) is equal to 1 means that the pixel Q belongs to character stroke and otherwise to background. 
Shown from Fig. 1 and formulation (2), (3), P;, P; and Ps, P; are used to detect vertical and horizontal stroke. 
  
306 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
- 
nt 
tc 
tc 
e 
F 
th 
Ci 
r( 
S 
F 
If 
th 
th