Académique Documents
Professionnel Documents
Culture Documents
A b s t r a c t : Thispaper presents a methodology for off-line handwritten Chinese characterrecognition based on mergence of
consecutive segments of adaptive duration. The handwritten Chinese character string is partitioned into a sequence of
consecutive segments, which are combined to implementdissimilarity evaluation within a sliding window whose durations
are determined adaptivelyby the integration of shapes and context of evaluations. The average stroke width is estimated for
the handwritten Chinese character string, and a set of candidate character segmentation boundaries is found by using the
integration of pixel and stroke features. The final decisions on segmentation and recognition are made under minimal
arithmetical mean dissimilarities. Experiments proved that the proposed approach of adaptive duration outperforms the
method of fixed duration, and is very effective for the recognition of overlapped, broken, touched, loosely configured
Chinese characters.
date segmentation boundaries are found, and the determined by searching the path of minimal ar-
candidate boundaries are confirmed by recognition ithmetical mean dissimilarity.
results. This method is more reasonable than the
first one for handwritten character recognition, but
it highly depends on the performance of the rec- ESTIMATION OF AVERAGE STROKE WIDTH
ognizer (Liang et al., 1994).
This paper's proposed recognition-based It is assumed that stroke width varies locally
methodology for handwritten Chinese character depending on writing device, paper texture and
segmentation and recognition makes the best use of pressure within a script. Therefore, it is reasonable
shape features and recognition results. For the to get average stroke width and use it for subse-
mergence of consecutive segments, a method of quent image processing procedures in adaptive
variable duration for each character based on sta- manner. We detect components of small size as
tistics for English characters is described in (Kim noise, positions of small stroke width as candidate
and Govindaraju, 1997). In this paper, the duration segmentation points, and shapes of some range of
for each character is determined adaptively by stroke width as components of characters.
shape and evaluation. To estimate the average stroke width of a
The set of candidate segmentation boundaries given handwritten character string (as shown in
is crucial to handwritten Chinese character recog- Fig.2a), MSW, contours of that image are extracted.
nition. We make the following assumption: the By tracing contours from the left-most column to
correct segmentation boundaries are embedded in the. right-most column, the following distances (as
the set of candidate segmentation boundaries. shown in Fig.2b) for each column are computed: (1)
In order to verify the proposed method, the distance between upper trace and lower trace of the
average stroke width is computed, then a set of outer contour, (2) distance between lower trace and
candidate segmentation boundaries is explored, and upper trace of the inner contour, (3) distance be-
segment-level evaluation based on combination of tween upper trace of inner contour and upper trace
consecutive components of adaptive duration is of the outer contour, and (4) distance between
described. A reduced directed graph is constructed lower trace of inner contour and lower trace of the
according to the sequence of evaluation results. The outer contour.
final segmentation and recognition decisions are The histogram in Fig.2c shows the number of
(a)
(a) (b)
(h)
600
(c) (d)
400
300
(e) (f)
I 200
,~176
0
1
,-.~.N
3
. . . . . .
5
(c)
7 9
N.m.~.m
11 13
occurrences of each distance value for the image of by the reduced nonlinear segmentation algorithm,
Fig.2b. The maximum of the histogram indicates an which split the connected characters or components.
estimation of the average stroke width. Fig.3b shows the candidate segmentation bounda-
We define a set of candidate segmentation ries from reduced nonlinear segmentation algo-
points rithm.
pm={no,nl . . . . ,riM} (1 1)
be correct if there exists a path from source to sink on of handwritings decreases the accuracy of
in the word graph. There are almost no cases not evaluation, which negatively affects the final rec-
completely partitioned in those text line images. ognition decision.
The final character segmentation and recognition
decisions are highly inter-related, and the correct
recognition accuracy of the top 5 candidate char- CONCLUSION
acters for this kind of script lines is 93%. The cor-
rect recognition accuracy of address words is We encountered difficulties from broken,
greater than 90%. overlapped, touched, and loosely configured char-
Because of the variety of character shapes acters for handwritten Chinese character segmen-
from handwriting styles, we make decision about tation and recognition. In this paper, we proposed a
segmentation and recognition based on the evalua- new strategy for handwritten Chinese character
tion of merged segments, instead of the charac- string segmentation and recognition that combines
terization of such character shapes as aspect ratio consecutive segments within a sliding window of
and width. The proposed strategy made a recogni- adaptive duration. The segmentation and recogni-
tion solution for not only touched and overlapped tion are highly inter-related. We have proved that
characters (as shown in Fig.5 and Fig.4), but also many of errors from these problems can be tackled
the combination of broken and loosely configured by the proposed approach.
characters (as shown in Fig.4). The methodology is There are still some ways to improve the
effective for segmentation and recognition decision segmentation and recognition: (1) exploiting con-
on the above characters, and greatly improved the text information from a lexicon, and (2) segmenting
recognition performance. the string into a sequence of finer segments will
Obviously, the proposed strategy of adaptive certainly improve the segmentation, while bringing
duration requires much less expensive computation about more expenses in computation.
resource than that of fixed duration, and improves
the performance of segmentation and recognition References
because of the optimization rule of minimal mean Favata, J.T., 2001. Offiine general handwritten word rec-
dissimilarity. ognition using an approximate BEAM matching algo-
rithm. I E E E Trans. on Pattern Recognition and Ma-
There are still main sources of segmentation chine Intelligence, 23(9): 1009-1021.
failure to be identified: (1) multi-touched character Kahan, S., Pavlidis, T., Baird, H.S., 1987. On the recogni-
string, (2) the accuracy of evaluations. The variati- tion of printed characters of any fonts and sizes. I E E E
Trans. on Pattern Recognition and Machine Intelli-
gence, 9(2):274-288.
Kim, G., Govindaraju, V., 1997. A lexicon driven approach
to handwritten word recognition for real-time appli-
cations. I E E E Trans. on Pattern Recognition and
(a) Machine Intelligence, 19(4):366-379.
Lee, S.W., Lee, D.J., 1996. A new methodology for
gray-scale character segmentation and recognition.
I E E E Trans. on Pattern Recognition and Machine In-
(b) telligence, 18(10): 1045-1050.
Liang, S., Shridhar, M., Ahmadi, M., 1994. Segmentation of
touching characters in printed document recognition.
Pattern Recognition, 27(6):825-840.
Nafiz, A., Fatos, T.Y., 2002. Optical character recognition
(c) for cursive handwriting. I E E E Trans. on Pattern
Fig.5 Experimental results Recognition and Machine Intelligence, 24(6):801-813.
(a) Text line image; (b) Candidate segmentation bounda-
ries; (c) Final segmentation and recognition result