You are on page 1of 23

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220412438

Chinese character recognition: History, status and prospects

Article  in  Frontiers of Computer Science in China · May 2007


DOI: 10.1007/s11704-007-0012-5 · Source: DBLP

CITATIONS READS

74 13,866

3 authors, including:

Cheng-Lin Liu Baihua Xiao


Institute of Automation of Chinese Academy of Sciences Chinese Academy of Sciences
318 PUBLICATIONS   10,551 CITATIONS    111 PUBLICATIONS   1,485 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Multi-Lingual Scene Text Detection and Recognition View project

All content following this page was uploaded by Cheng-Lin Liu on 10 September 2014.

The user has requested enhancement of the downloaded file.


Chinese Character Recognition: History, Status, and
Prospects

DAI Ruwei1, LIU Chenglin2, and XIAO Baihua1


1 Laboratory of Complex Systems and Intelligence Science, Institute of Automation, Chi-
nese Academy of Sciences, Beijing 10080, China
{ruwei.dai,baihua.xiao}@ia.ac.cn
2 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of

Sciences, Beijing 10080, China


liucl@nlpr.ia.ac.cn

Abstract. Chinese character recognition (CCR) is an important branch of pat-


tern recognition. It was considered as an extremely difficult problem due to the
very large number of categories, complicated structures, similarity between
characters, and the variability of fonts or writing styles. Because of its unique
technical challenges and great social needs, the last four decades witnessed the
intensive research in this field and a rapid increase of successful applications.
However, higher recognition performance is continuously needed to improve
the existing applications and to exploit new applications. This paper first pro-
vides an overview of Chinese character recognition and the properties of Chi-
nese characters. Some important methods and successful results in the history
of Chinese character recognition are then summarized. As for classification
methods, this article pays special attention to the syntactic-semantic approach
for online Chinese character recognition, as well as the meta-synthesis ap-
proach for discipline crossing. Finally, the remaining problems and the possible
solutions to them are discussed.

1 Introduction

Chinese character recognition is an important branch of pattern recognition [1-5]. The


solution of this problem relies on many techniques in various fields: image processing,
machine learning, cognitive science (noetic science), linguistics, etc. From the start of
pattern recognition research in 1950s, character recognition has been a major test case
and a stimulator of pattern recognition methodology. At the first workshop on pattern
recognition, held in Puerto Rico, US, 1966, about one third of papers were dealing
with character recognition [6]. Approaches such as blurring [7], directional pattern
matching [8,9], hierarchical classification [10,11] and multiple classifiers combina-
tion [12,13] were first proposed by the character recognition community, and later
evolved into attractive research fields.
Character recognition systems contribute tremendously to the advance of the auto-
mation process and can be of significant benefit to man-machine communication in
many applications, such as postal mail sorting, business card reading, bank checks

1
and transaction forms processing, and recently, in digital libraries and mobile phones.
Chinese characters are used by over 1.3 billion people in China and some other coun-
tries or areas, but typing Chinese characters into computers is not a trivial tasks. In
China, many people cannot even use phonetic codes for character entry because they
habitually speak dialect and cannot pronounce Mandarin correctly. So, the automatic
recognition of Chinese characters would have widespread special benefits.
Chinese character recognition was considered as an extremely difficult problem
due to the very large number of categories, complicated structures, similarity between
characters, and the variability of fonts or writing styles. Due to its unique technical
challenges and great social needs, the last four decades witnessed the intensive re-
search in this field and a rapid increase of successful applications. This paper pro-
vides a brief review of this field, outlines the important methods and advances, and
discuss the potential future research directions.
The approaches of character recognition is dichotomized into online and offline
depending on the hardware and application mode. It is called “online” if the temporal
sequence of pen trajectory (captured by, e.g., digitizing tablet) is available. The pen
trajectory is immediately recognized after it is written, and the user can respond to the
recognition result (to correct the result or re-write). It is “offline” if to recognize pre-
viously written text, which is converted to images using a scanner or a camera. This
paper covers both online and offline Chinese character recognition.
The rest of this paper is organized as follows: Section 2 describes the properties of
Chinese characters, Section 3 briefly reviews the history of Chinese character recog-
nition and the state of the art. Section 4 addresses from syntactic to syntactic-semantic
approach and its applications to online Chinese Characters Recognition. Section 5
discusses the discipline crossing between pattern recognition and systems science, as
well as the resulting meta-synthesis approaches. Finally, section 6 discusses the re-
maining problems and the possibilities for solving them.

2 Properties of Chinese Characters

Chinese characters have unique structures compared to western characters and this
uniqueness poses technical challenges to recognition. This section summarizes the
properties of Chinese characters as follows.

2.1 Evolution of Chinese Characters

Fig. 1 demonstrates the evolution of Chinese characters. The origin of Chinese char-
acters can be traced back to oracle script and script on bronze before 1000 BC. Offi-
cial script was invented in Qin Dyanisty (about 220BC), and got popular in Han Dy-
nasty. Its shape is very similar to the contemporary characters. Regular script, cursive
script and fluent script were invented in late Han Dynasty (about 180AD). After that
time, while spoken Chinese varies across regions, the written Chinese characters
remain relatively stable. The regular script, cursive script and fluent script have been
commonly used until today. However, as you can see in Fig.2, the traditional Chinese

2
characters have too many strokes. To ease writing, Chinese government carried out
Chinese character reformation and published 2,235 simplified characters during
1956-1964. The average number of strokes for the 2,235 characters was reduced from
16.03 to 10.3. The simplified Chinese characters together with the characters that
were not simplified, come to be a standard for official communication across China.

Fig. 1. Examples of the evolution of Chinese characters (from left to right: ‘sun’, ‘moon’,
‘vehicle’, and ‘horse’).

Fig. 2. Examples of traditional and simplified Chinese characters (upper for traditional, lower
for simplified).

2.2 Chinese Character Set

Chinese characters are used in daily communications by over one quarter of world’s
population, mainly in Asia, such as China, Korea, Japan, and Singapore. There are
mainly three character sets: traditional Chinese characters, simplified Chinese charac-
ters, and Japanese Kanji [5].
In Japan, 2,965 Kanji characters are included in the JIS level-1 standard and 3,390
Kanji characters are in the level-2 standard. Japanese Kanji characters have mostly
identical shape to the corresponding traditional Chinese or simplified Chinese.
In Taiwan of China, 5,401 traditional characters are included in a standard set. In
the mainland of China, three character sets, containing 6,763, 20,902 and 27,533
Chinese characters, respectively, were announced as the National Standards (see
Table 1). The 6,763 characters in GB2312-80 covers 99.99% of usage, but still do not
suffice. Especially, many characters used in human names and place names are not
included in this set. A general-purpose recognizer needs to cover about 9,000 simpli-
fied characters, about 3,000 of which have different traditional shapes. In addition,
about 1,000 symbols and special characters should be included. In experiments of
academic research, usually 3,755 characters are considered. The very large number of

3
categories poses a technical challenge for efficient and accurate classification of Chi-
nese characters.

Table 1. National standards of Chinese character set.


National standard Number of characters Description
GB2312-80 Level-1: 3,755 Simplified
Level-2: 3,008
(Totally 6,763)
GBK 20,902 Simplified and Traditional
GB18030-2000 27,533 Plus characters of minority
nationalities

2.3 Character Structures

Chinese characters are ideographs with complicated structures. Many Chinese charac-
ters contain relatively independent substructures, called radicals, and some common
radicals are shared by different characters. That is to say, a Chinese character is com-
posed of radicals, which are in turn composed of straight-line or poly-line strokes (see
Fig. 3).
As far as we know, the most complicated Chinese character has 36 strokes, see the
bottom left of Fig. 3. The total number of radicals and single-component characters in
Chinese characters is about 500.

Fig. 3. Examples of Chinese character structures (the right panel shows a complicated Chinese
character with five radicals, some of which can be further decomposed).

The pattern of Chinese character structures can be roughly categorized into 10


types (single-radical, left-right, up-down, up-right, left-down, up-left-down, left-up-
right, left-down-right, and enclosure), see Fig. 4. Some of the patterns can be further
divided into sub-categories.
The structural complexity of Chinese characters is a merit for recognition: it carries
rich information for discriminating different characters. This hierarchical character-
radical-stroke structure can be utilized in recognition to largely reduce the size of
reference model database and speed up recognition. However, the complexity of
structures makes the structural description difficult.

4
Fig. 4. 10 types of Chinese character structures (single-radical, left-right, up-down, up-left, up-
right, left-down, up-left-down, left-up-right, left-down-right and enclosure).

Besides the large number of categories and the complexity of structures, there are
many similar Chinese characters which differ only slightly (see Fig. 5). The similar
characters are hard to discriminate by computer recognizers.

Fig. 5. Examples of similar Chinese character pairs.

2.4 Writing Styles

The enormous writing styles of different persons can be roughly divided into three
categories: regular script (also called handprint), fluent script, and cursive script. The
intermediate style between regular and fluent is called fluent-regular, and the inter-
mediate between fluent and cursive is called fluent-cursive [5]. Some examples of the
three typical styles are shown in Fig. 6. We can see that the strokes of regular script
are mostly straight-line segments. The fluent script has many curved strokes and,
frequently, successive strokes are connected. In cursive script, some character shapes
differ drastically from the standard shape.

5
Fig. 6. Examples of three major writing styles (from left to right: regular script, fluent script,
cursive script).

3 Historical Review of the Technology

In the following, we review the history of Chinese character recognition in respect of


the evolution of recognition target, the evolution of recognition methods, the citation
of important techniques and representative results, and finally, we summarize the
recent advances in the state of the art.

3.1 Evolution of Recognition Target

Since Casey and Nagy published the first work of printed Chinese character recogni-
tion [10], the recognition of Chinese characters has evolved into an attractive research
area. The evolution of recognition target, from machine-printed to handwritten, from
online to offline, basically observes an easy-to-difficult order.
As summarized in Table 2, both printed Chinese character recognition and online
handprinted Chinese character recognition were started in mid-1960s, by researchers
in US. Researchers in Japan began to deal with printed and online handprinted Chi-
nese (Kanji) character recognition in late 1960s, and offline handprinted Chinese
character recognition in late 1970s. Chinese researchers started area about 10 years
later than those in Japan, i.e., printed and online handprinted Chinese character rec-
ognition in late 1970s, and offline handprinted Chinese character recognition in late
1980s. From 1990s, handprinted Chinese character recognition (especially online
recognition) has been well commercialized, and the research target has moved to less-
constrained handwritten character recognition.

Table 2. Evolution of recognition targets.


Time Recognition target
Mid-1960s~ Printed Chinese (IBM) [10]
Online handprint (MIT, U. Pittsburgh) [14,15]
Late 1960s~ Printed Chinese, online handprint (Japan)
Late 1970s~ Offline handprint (Japan)
Printed Chinese, online handprint (China)
Late 1980s~ Offline handprint (China)
Mid-1990s~ Less-constrained handwritten

3.2 Evolution of Recognition Methods

From the 1960s, many effective methods have been proposed in the area of Chinese
character recognition. The evolution of major methods is summarized in Table 2.
Template matching, including one-stage classification and hierarchical classification,
was widely used in early works of Chinese character recognition, especially printed

6
character recognition. Character structure analysis (stroke analysis, relaxation match-
ing, attributed graph matching) attracted much attention during 1970s-1990s. Espe-
cially, structural matching using relaxation and attributed graphs was popular in
1980s-1990s. Following template matching and pattern matching, feature matching
got popular in 1980s [2]. It provides good feature extraction techniques for the cur-
rent statistical classification methods. From 1990s, statistical recognition methods
dominate the technology. However, structural methods are still under study, because
they resemble the procedure of human cognition and have the potential of recogniz-
ing cursively handwritten characters. A recent advance is the statistical modeling of
character structures [21].

Table 3. Evolution of major recognition methods.


Time Recognition method
1960s-1970s Hierarchical template matching [10,11]
1970s-1990s Structural (stroke analysis, relaxation, attributed graph)
[1][16], Syntactic (attributed grammar) [17-19]
1980s~ Feature matching [2],
Statistical classification [20]
1990s~ Statistical methods dominate,
Statistical structure modeling [21]

3.3 Important Techniques

In addition to the evolution of general methods as reviewed in the last section, this
section summarizes the important detailed techniques that have affected the technol-
ogy of this field, i.e., they have yielded satisfactory recognition results or made appli-
cations successful. Table 4 gives a list of such important techniques, including those
for pre-processing, feature extraction, classification, etc. Most of the techniques are
still actively used now.
The first paper on printed Chinese character recognition, by Casey and Nagy in
1966, propose the technique of hierarchical template matching [10]. This is the origin
of current multi-stage classification, which is efficient to speed up the classification
of large category set. Blurring was proposed by Iijima in 1960s, and got internation-
ally know when it was published in 1973 at the 1st International Joint Conference on
Pattern Recognition (IJCPR) [7]. Blurring is equivalent to the current spatial filtering,
and is effective to reduce image noise and improve the translation invariance. Direc-
tional pattern matching was first published in 1979 by Yasuda and Fujisawa [8], and
got widely known after a paper was published in 1983 at Pattern Recognition Letters
(Yamashita et al. [9]). Directional pattern matching is the origin of the popularly used
direction feature extraction, and is often used together with blurring.
The work of Yamamoto and Rosenfeld on Chinese character recognition using re-
laxation matching [16], started a boom of relaxation-based structural matching in
1980s and 1990s. Nonlinear normalization based on line density equalization is effec-
tive to reduce the within-class shape variation, and thus greatly improve the recogni-
tion accuracy. It was first published in 1984 in Japanese, and internationally at ICPR
1988 [22] and Pattern Recognition Journal in 1990 [23].

7
Before the modified quadratic discriminant function (MQDF) of Kimura et al. was
proposed, Chinese character recognition were mostly performed by a simple distance-
based classifier. The MQDF, with lower complexity and higher generalization per-
formance than the ordinary QDF, is demonstrated superior in handwritten Chinese
character recognition. It was first published in ICPR 1984, and got widely known
after publication in IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI)
in 1987 [20].
Decision-tree classification, intensively studied by Suen’s group in Concordia Uni-
versity [24,25], is superior in classification speed for large category set, and has ap-
plied successfully to printed Chinese character recognition
The semantic-syntactic approach, proposed in 1980s by J.W. Tai (R.W. Dai)
[19,26], has largely affected the technology of online Chinese character recognition.
Though multiple-classifier approaches have been intensively studies from the begin-
ning of 1990s, it was not successfully applied to large category set problems until the
researchers from Institute of Automation, Chinese Academy of Sciences (CASIA)
and Tsinghua University published their works of Chinese character recognition by
combining multiple classifiers [27,28].

Table 4. Important techniques in Chinese character recognition.


Year Technique Authors
1966 Hierarchical template matching [10] Casey & Nagy (IBM)
1960s Blurring (spatial filtering) [7] T. Iijima (TIT, Japan)
1979 Directional pattern matching [8,9] Fujisawa & Yasuda
(Hitachi),
1982 Relaxation matching [16] K. Yamamoto
(ETL)
1984 Nonlinear normalization [22,23] Yamada (ETL)
Tsukumo (NEC)
1984 Modified QDF [20] Kimura et al. (Mie U)
Mid-1980s Decision tree classification [24,25] Suen’s group (Concordia)
Mid-1980s Semantic-syntactic (attributed grammar) J.W. Tai (R.W. Dai),
[19,26] Y.J. Liu (CASIA)
1997 Multiple classifiers fusion [27,28] R.W. Dai (CASIA),
X.Q. Ding (Tsinghua)

3.3 Important Results

Many experiments of Chinese character recognition, mostly using the effective tech-
niques listed in Table 4, have reported high performance. To evaluate the perform-
ance, the ETL8B and ETL9B databases, collected by Electro-Technical Laboratory of
Japan, have been widely tested. The ETL8B database contains the handprinted im-
ages of 952 characters (881 Kanji and 71 hiragana), 160 samples per class. The
ETL9B database contains the handprinted images of 3,036 classes (2,965 Kanji and
71 hiragana), 200 samples per class. In China, some experiments have been con-
ducted on the HCL2000 database (collected by Beijing University of Posts and Tele-
communications) and the CASIA database (Institute of Automation, Chinese Acad-

8
emy of Sciences), both for 3,755 Chinese characters, with 1,000 samples per class
and 300 samples per class, respectively.
Some representative results of online Chinese character recognition have been
summarized in [5]. In this paper, we give a list of representative results of offline
Chinese character recognition in Table 5. For reference, the underlying normaliza-
tion/feature extraction and classification methods are also given. “NLN” denotes
nonlinear normalization, which varies slightly in implementation.

Table 5. Important results in Chinese character recognition.


Year Author Feature Classifier Data- Accu-
base/class racy
1983 Yamashita [9] Direction Correlation ETL8/881 94.8%
1984 Yamamoto [29] Contour Relaxation ETL8/952 99.0%
1984 Yamada [30] Contour DP ETL8/952 99.6%
1986 Yamamoto [31] Contour Relaxation ETL9/3036 98.5%
1988 Tsukumo [22] NLN/Direction Correlation ETL9/3036 94.42%
1992 Tsukumo [32] NLN/Direction Flexible match ETL9/3036 95.05%
1993 Jun Guo [33] NLN/Direction Perturbation ETL9/3036 96.32%
1996 Saruta [34] NLN/Direction Neural net ETL9/3036 95.48%
1997 Suzuki [35] NLN/Direction Mahalanobis+ ETL9/3036 99.31%
1997 Kimura [36] NLN/Direction MQDF ETL9/3036 99.15%
1999 Kato [37] NLN/Direction Mahalanobis ETL9/3036 99.42%
2001 Sawa [38] NLN/Gradient MQDF+Perturb ETL9/3036 99.41%
2003 J.X. Dong [39] NLN/Gradient SVM ETL9/3036 99.0%
2005 H. Liu [40] NLN/Gradient MQDF+MCE+ HCL/3755 98.56%
2006 C.L. Liu [41] NLN/Gradient DFE-MQDF CASIA/3755 98.43%
Yamashita et al. first applied directional pattern matching to a relatively large char-
acter set and reported a high accuracy on a test set of 20 samples per class of ETL8B
[9]. The methods of Yamamoto et al. [29], Yamada [30], and Yamamoto et al [31] are
basically contour segment matching. They reported very high accuracies on small test
sets of only a few samples per class.
Among the experiments on ETL9B database, Tsukumo and Tanaka [22], Tsukumo
[32], Guo et al [33], and Saruta et al. [34] used 100 samples per class for training
classifiers and the disjoint 100 samples per class for testing. All the four works use
nonlinear normalization and contour direction feature, while the classification meth-
ods are correlation (simple distance), flexible pattern matching, perturbation-based
correlation, and class-module neural network, respectively.
From 1997, the experiments on ETL9B database mostly use 160 or more samples
per class for training and the remaining samples for testing. They all use nonlinear
normalization and contour or gradient direction feature, with finer implementations.
The higher accuracies are also partially due to the larger number of training samples.
Suzuki et al. use a modified Mahalanobis distance for classification and an auxiliary
measure for pairwise discrimination [35]. Kimura et al. use a pseudo Bayes classifier,
which is similar to MQDF, for classification [36]. The classifier of Kato et al. is an
asymmetric Mahalanobis distance function [37]. Sawa et al. generated deformed
training samples for estimating the parameters of MQDF [38]. Dong et al. proposed a

9
fast algorithm for training support vector machine (SVM) classifiers on large data set
and applied successfully to Chinese character recognition [39].
H. Liu and X. Ding (Tsinghua University) [40] and C.L. Liu (CASIA) [41] ex-
perimented on ETL9B database as well as a Chinese database. They both use nonlin-
ear normalization and gradient direction feature, and for classification, they combine
MQDF with minimum classification error (MCE) training or discriminative feature
extraction (DFE). Their accuracies on the test samples of ETL9B are over 99.3%, and
the accuracies on HCL2000 and CASIA database are 98.56% and 98.43%, respec-
tively. This indicates that the characters written by Chinese are more difficult to rec-
ognize than those written by Japanese.

3.4 The State of the Art

In the following, we discuss generally the status of technology and highlight some
recent advances in respect of pre-processing, feature extraction, feature transforma-
tion, and classifier design.
3.4.1 Character Pre-Processing
The main pre-processing steps including noise reduction (generally by smoothing or
low-pass spatial filtering) and character shape normalization. Normalization is more
influential to the recognition performance. It not only standardizes the image size, but
also reduces the within-class variation of character shape.
Nonlinear normalization based on line density equalization [22,23], has contrib-
uted significantly to the improvement of performance in handwritten Chinese charac-
ter recognition. A pseudo two-dimensional (P2D) nonlinear normalization method
through line density smoothing [42], can further improve the recognition accuracy.
This method, however, is very time consuming. Recently, a new P2D normalization
method based on line density projection interpolation was proposed by Liu et al. [43].
It performs comparably well with the method of [42] with only a little extra complex-
ity than 1D normalization. Compared to 1D nonlinear normalization, P2D normaliza-
tion can not only correct the non-uniform stroke density, but also alleviate the imbal-
ance of width/height and stroke positions in different parts of the character image.
3.4.2 Feature Extraction
For feature extraction, the features are hoped to describe more details of the character
shape, be invariant against within-class shape variation, and reflect between-class
difference. Better tradeoff between the within-class invariance and the between-class
discrimination can be achieved by structural element decomposition (e.g., local stroke
direction decomposition) and image (or feature map) blurring.
Early works usually used 4-orientation decomposition to extract direction features.
8-direction decomposition is found to give higher recognition accuracy. Possibly, it
can be extended to 12-direction and 16-direction. In the case of 8-direction contour
(chaincode) or gradient direction decomposition, the two sides of stroke edge are
treated in different directions, such that the confusion between parallel strokes can be
better differentiated. The local gradient direction of character image can be parti-
tioned into a number of angle ranges [44]. A gradient vector decomposition approach,

10
originally proposed in online character recognition [45], has yielded superior recogni-
tion performance in handwritten character recognition [46].
Comparing chaincode direction feature and gradient direction feature, both are in-
sensitive to the stroke-width variation and yield high recognition accuracies in Chi-
nese character recognition, but chaincode feature applies to binary image only, while
gradient feature applies to gray-scale image as well. The gradient feature generally
outperforms the chaincode feature because it is more stable against image noise and
the fluctuation of local contour direction. Nevertheless, the computation of gradient
feature is a little more complicated than that of chaincode feature.
Extracting gradient direction feature directly from gray-scale images will be trend
in the future, especially for low-resolution or degraded character images.
3.4.3 Feature Transformation
After feature extraction, the reduction of dimensionality is important for both reduc-
ing the computation complexity of classification and improving the generalization
accuracy. The dimensionality of the feature vector of direction feature, for example,
is as high as 512 (8-direction, 8x8 sampled values for each direction) or more. When
using a nonlinear classifier like the MQDF directly on this high-dimensional vector,
the storage and computation complexity will be very high. Hence, dimensionality
reduction is now widely adopted in character recognition.
Dimensionality reduction is performed by projecting the feature vector onto a low-
dimensional linear subspace. The most popular linear dimensionality reduction tech-
niques are the principal component analysis (PCA) and the linear discriminant analy-
sis (LDA) [47]. Unlike the PCA that maximize the variance of data vectors in sub-
space regardless of the class labels of vectors, the LDA learns a subspace that maxi-
mize the ratio of between-class scatter to within-class scatter. LDA has shown prom-
ise in Chinese character recognition [36].
Though performs fairly well in practice, LDA has some inherent drawbacks. It as-
sumes equal-covariance Gaussian densities for all classes, and does not separate well
nearby classes in subspace. The heteroscedastic discriminant analysis (HDA, e.g.,
[48]) considers the difference of covariance in subspace learning, but is extremely
expensive for large category set. Very recently, H. Liu and X. Ding proposed a new
HDA method [40], which is computationally feasible for large category set and yields
higher accuracy than LDA. Discriminative feature extraction (DFE), which adjusts
the subspace axes with the aim of minimizing the classification error on training data,
has yielded significant improvement of accuracy in handwritten Chinese character
recognition [41].
3.4.4 Classifier Design
As to classifier design, there is a tradeoff between the classifier complexity and the
classification accuracy. Typical classifiers include: (1) Minimum distance (or correla-
tion) classifier, which was widely used before 1980s; (2) LVQ (learning vector quan-
tization) for prototype optimization, which offers good tradeoff between complexity
and accuracy [49]; (3) Modified quadratic discriminant function (MQDF, proposed
by Kimura et al in 1980s [20]), which gives high accuracy but involves a large num-
ber of classifier parameters.

11
To speed up the classification of large category set, hierarchical classification has
been commonly adopted from the early work of Chinese character recognition.
Suen’s group reported progress in Chinese character recognition using decision tree
classifiers in 1980s [24,25]. More often, candidate reduction by class grouping (clus-
tering) or multi-stage dynamic candidate selection are adopted. Typical works using
both two schemes include the printed Chinese character system of Hitachi [11] (pub-
lished in 1st IJCPR 1973), and the work of handprinted Chinese character recognition
by Y.Y. Tang et al [50].
More recently, multiple classifiers combination approaches have attracted more
and more interests. Combination of multiple classifiers is effective to improve the
recognition accuracy of single classifiers. Suen’s group in Concordia University did
many works on combining classifiers for small character set recognition (e.g., hand-
written numerals) [12]. Combining classifiers for large character set faces some prob-
lems: simple majority vote does not perform sufficiently, while the training of com-
biner on large data set is not trivial. Researchers from CASIA and Tsinghua Univer-
sity reported success of multiple classifiers in Chinese character recognition: one uses
weighted confidence fusion [28] and the others are based on meta-synthesis [27,51-
55].

4 From Syntactic to Syntactic-Semantic Approach

Generally speaking, the mathematical methods for solving pattern recognition prob-
lems can be grouped into two major categories: statistical (or decision theoretic) ap-
proach and syntactic (or structural) approach. Statistical pattern recognition is based
on statistical characterizations of patterns, assuming that the patterns are generated by
a probabilistic system. Structural pattern recognition is based on the structural interre-
lationships of features. The former approach has been intensively studied and widely
used. This section will focus on the progress of the latter approach, from syntactic to
syntactic-semantic.

4.1 Syntactic Pattern Recognition

The structure information has been regarded as very important in pattern analysis. It
is well known that Prof. K.S. Fu is the pioneer of the syntactic pattern recognition. He
has published many papers and books in this field [17,18,56-64]. This approach
draws an analogy between the structure of patterns and the syntax of languages, see
Fig. 7.

12
Fig. 7. Block diagram of a syntactic pattern recognition system [17].
Different kinds of grammars have been proposed to describe patterns. It is hoped
that in order to describe a class of patterns,the grammar used can be directly in-
ferred from a set of sample strings or a set of sample patterns. The problem of learn-
ing a grammar based on a set of sample strings is called grammatical inference. Actu-
ally,even some simple two-dimensional patterns(such as equilateral triangles) have
to be described by context-sensitive languages. However,the inference of general
context-sensitive grammar may not be desired since recognition of context-sensitive
languages is in general rather complex and time consuming.

4.2 Syntactic-Semantic Approach

The syntactic approach was considered to expand in 80s last century. Strings, trees,
and graphs have been suggested to describe pattern structures. The basic idea is to
represent a pattern by its components (sub-patterns) and the relations between them.
By using relational graphs to represent pattern structures, the simple “concatenation”
relation can be extended to include many other relations. And semantic information in
pattern description can be included by adding attributes to the description of sub-
pattern and relations. Besides, the description and recognition of patterns by means of
attributed grammars have been advocated by several investigators [17,59-64].
Attributed grammars were first formulated by Knuth [65]. And an attributed gram-
mar method for pictorial pattern recognition was developed in China [19,26,66], and
successfully applied for solving online Chinese character recognition problem. The
attributed grammar including two parts was defined as follows:
(1) A syntactic part represented by a context-free or finite-state grammar.
(2) A semantic part consisting of three sets
1) A set of primitive attributes
2) A set of relational attributes
3) A set of semantic functions or semantic rules
Such an “attributed grammar approach” was termed as “syntactic-semantic ap-
proach”.
For the attributed grammar,there is a trade-off between syntactic and semantic
complexities in grammar definition, see Figure 8. That is, semantic information can
be applied to achieve lower syntactic complexity in pattern description. Thus, attrib-
uted finite-state instead of context-sensitive languages, can be used as a normal form

13
for pattern description. One could find that both the traditional statistical and syntac-
tic approaches are the special cases of syntactic-semantic approach.

Fig. 8. Syntactic-semantic approach: tradeoff between syntactic and semantic.

5 Discipline Crossing

Character recognition is highly related to various fields: image processing, machine


learning, cognitive science (noetic science), linguistics, etc. Meta-synthesis was pro-
posed by Chinese scientists in 1990 for dealing with open complex giant systems
(OCGS) [67]. This section will briefly introduce the progress on discipline crossing
between pattern recognition and systems science, and the related meta-synthetic ap-
proach for pattern recognition system design.

5.1 An Introduction of OCGS and Meta-synthesis

To facilitate research on systems science,systems can be divided using different


principles into different class types. Depending on the number and variety of subsys-
tems contained in the system, and the number of interactions between them,
systems can be divided into two groups: simple systems and giant systems. If there
are a large variety of subsystems with a hierarchical structure; and the interaction
between basic unit and basic unit, basic unit and environment is represented by some
mathematical formula,such as nonlinear function, or by an information protocol,
then the aggregate is called a complex giant system. In addition, if a system and its
subsystems exchange energy, information or material with environment, it is called an
open system.
The method dealing with OCGS is beyond the reductionism, studies and practice
have clearly proven that the only feasible and effective way to treat an OCGS is a
meta-synthesis from qualitative to the quantitative,i.e. the meta-synthetic engineer-
ing method. The main ideas of meta-synthesis are summarized as follows: human
computer cooperation, integration and system point of view.

5.2 Meta-synthesis Approach for Pattern Recognition Systems Design

In order to show the meta-synthesis related to pattern recognition, an example of


traditional Chinese medical diagnosis is given in Figure 9. In traditional Chinese
medical treatment, the medical doctor makes acquisition information from patient by
means of looking, hearing, asking and feeling the pulse. These pieces of information

14
are put into the doctor's brain, and according to his (or her) medical experiences, he
makes meta-synthesis, then gives the result of diagnoses.

Fig. 9. An example of meta-synthesis (Traditional Chinese medical treatment, from the situa-
tion of human body to consider the synthetic diagnosis).
Enlightened by the ideas of meta-synthesis mentioned above,the meta-synthetic
approaches were proposed in China as a guided framework to deal with complicated
pattern recognition system design. The characteristics of the meta-synthetic approach
for pattern recognition system design can be summarized as follows : Human-
Computer Cooperation,Integration and Closed-loop system. Fig. 10 demonstrates
the framework of such approaches.

Fig. 10. Block diagram of a pattern recognition system using meta-synthesis approach.

5.3 Multiple Classifiers Combination for CCR

Multiple classifiers integration has received considerable attention in the past decade.
The idea appeared under many names: hybrid methods, classifier combination, infor-
mation fusion, ensemble learning, etc. In meta-synthetic approach, the integration
includes not only multiple classifiers combination, but also human-computer integra-
tion by supervised learning.
As mentioned above, Chinese character recognition was considered as an ex-
tremely difficult problem due to the very large categories, complicated structures,
similarity between characters, and variability of fonts or writing styles. The large
categories characteristic is the main obstacle that prevents many supervised learning

15
algorithms (such as artificial neural networks, etc.) from successful application in
CCR.
To solve this problem, some researchers started research on multiple classifiers
combination using the ideas of meta-synthesis from 1990s, and reported promising
higher performance on handwritten Chinese character recognition [27,51-55]. In 1997,
Hao proposed an initial meta-synthesis approach for handwritten Chinese character
recognition, which used linear subnet to train the integration network [35]. As is
known, the human is proficient at judging which class the training sample belongs to,
instead of judging the suitable weights directly. Thus, an adaptive weighted multiple
classifiers combination method was proposed, where supervised learning can be ap-
plied indirectly [53]. In order not only to improve the performance of pattern recogni-
tion system, but also to pursue feasible human-machine integration scheme, a new
parallel compact integration scheme based on multi-layer perceptron (MLP) networks
is proposed by Wang [54] to solve handwritten Chinese character recognition prob-
lem, see Figure 11. In this approach, MLP network classification and integration can
be applied reasonably and effectively to solving large vocabulary classification prob-
lem.

Fig. 11. Totally parallel integration with two-step supervised learning.


Figure.11 shows the parallel compact integration system, where L compact MLP
network classifiers are integrated together to solve an N-category classification prob-
lem. Outputs of all compact MLP network classifiers are combined together to be an
enhanced feature, which is inputted to the integration network. Each output node of
the integration MLP network corresponds to a category, and its output value is in

16
direct proportion to the similarity between the input pattern and its corresponding
category.
Effective human-machine integration is realized through the procedure of two-step
supervised learning, which takes full advantage of the intelligence of teacher. In the
first step, the compact MLP network classifiers are constructed and trained, and then
the integration network is constructed and trained in the second step. During the su-
pervised learning procedure, the relationship among different categories is learned
and recorded in the compact MLP network classifiers, and the relationship among
different compact MLP network classifiers is learned and recorded in the integration
network too.

5.4 Pattern Recognition with Feedback

Furthermore,the meta-synthetic ideas can be applied to pattern recognition system


design to change the traditional viewpoint, which leads to the incorporating feedback
into classifiers integration network.
The integrated pattern recognition system incorporating feedback is no longer a
traditional nonlinear forward mapping, but a closed-loop nonlinear dynamical system.
The importance of this point view is that denotes a connection between pattern recog-
nition field and control system field.
This method has been successfully applied to handwritten numeral recognition
[53], see Fig. 12.

17
Fig. 12. Integration network with feedback.

6 Future Directions

The gap between the technical status and the required performance indicates that the
problem of Chinese character recognition is not solved completely yet and it leaves us
research opportunities. For examples, on three sets of handwritten Chinese character
samples (regular script, fluent script, and cursive script) shown in Fig. 6, a state-of-
the-art recognizer gives accuracies of 98%, 82%, and 70%, respectively. The per-
formance on fluent script and cursive script are far from satisfaction. To improve the
recognition performance, Table 6 gives a list of the remaining technical problems and
potential solutions.

Table 6. Remaining problems and potential solutions.


Issue Problem Solution
Database Insufficient training data, esp. Collection of new data, artifi-
(very important) fluent and cursive writing cial samples
Normalization Tradeoff between shape resto- New normalization algorithms,
ration and distortion perturbation
Feature extraction Features for discriminating Higher-order feature detectors,
similar characters feature selection
Classification Insufficient accuracy, esp. Learning from large data, multi
similar characters and cursive heterogeneous classifiers
Structural matching Model building, stroke extrac- Structural learning, model-
tion based stroke extraction
Character segmenta- Character splitting and merg- Character detection relying on
tion ing robust classification
Contextual processing Statistics (n-gram): not suffi- Syntactic/semantic text analysis
cient

The first problem is the lacking of large sample database for training classifiers.
Especially, the samples of fluent/cursive handwriting are not sufficient.
For statistical classification, the techniques of normalization, feature extraction,
and classification can be further improved: (1) Current normalization methods trade-
off between within-class variation reduction and shape distortion. New normalization
methods are to be proposed, or perturbation method can work well. (2) In feature
extraction, we can design higher-order feature detectors for extracting more discrimi-
native features, but need feature selection to overcome the curse of dimensionality. (3)
Learning classifiers from large sample data has not paid enough attention in Chinese
character recognition. Some classifiers like neural networks and support vector ma-
chines (SVMs), encounters difficulty when applying to large category set classifica-
tion problems.
Structural matching needs efficient algorithms for stroke extraction and structural
model learning. Syntactic-semantic methods need to be further studied, because they

18
resemble the procedure of human cognition and have the potential of recognizing
cursive script Chinese characters.
In addition, for practical applications, character segmentation and contextual proc-
essing should be paid high attention. Segmentation is a character detection problem,
good classifiers can help. Contextual processing should exploit more syntac-
tic/semantic knowledge.

Acknowledgements

Some parts of this paper have been presented at the 18th ICPR in Hong Kong, Au-
gust 2006, as a keynote speech. The authors would like to thank Prof. Yuan Yan
Tang, Prof. Xiaoqing Ding and Prof. Chunheng Wang for providing materials to
assist with this paper, and the reviewers for suggestions of improvements. Also, the
authors would like to apologize to researchers whose works are overlooked.

References

1. W. Stallings, Approaches to Chinese character recognition, Pattern Recognition, 8(2): 87-


98, 1976.
2. S. Mori, K. Yamamoto, M. Yasuda, Research on machine recognition of handprinted
characters, IEEE Trans. Pattern Analysis and Machine Intelligence, 6(4): 386-405, 1984.
3. M. Umeda, Advances in recognition methods for handwritten Kanji characters, IEICE
Trans. Information and Systems, E29(5): 401-410, 1996.
4. T.H. Hildebrandt, W. Liu, Optical Recognition of Handwritten Chinese Characters:
Advances since 1980, Pattern Recognition, Vol. 26, No. 2, pp. 205-225, 1993.
5. C.-L. Liu, S. Jaeger, M. Nakagawa, Online recognition of Chinese characters: the state-
of-the-art, IEEE Trans. Pattern Analysis and Machine Intelligence, 26(2): 198-213, 2004.
6. G. Nagy, Pattern Recognition 1966 IEEE Workshop, IEEE Spectrum, Feb. 1967, pp.92-
94.
7. T. Iijima, H. Genchi, K. Mori, A theory of character recognition by pattern matching
method, Proc. 1st IJCPR, 1973, pp.50-56.
8. M. Yasuda, H. Fujisawa, An improved correlation method for character recognition,
Systems, Computers, and Controls, 10(2): 29-38, 1979 (Translated from Trans. IEICE
Japan, 62-D(3): 217-224, 1979).
9. Y. Yamashita, K. Higuchi, Y. Yamada, Y. Haga, Classification of handprinted Kanji
characters by the structured segment matching method, Pattern Recognition Letters, 1:
475-479, 1983.
10. R. Casey, G. Nagy, Recognition of printed Chinese characters, IEEE Trans. Electronic
Computers, EC-15(1): 91-101, 1966.
11. S. Yamamoto, A. Nakajima, K. Nakata, Chinese character recognition by hierarchical
pattern matching, Proc. 1st IJCPR, 1973, pp.183-194.
12. L. Xu, A. Krzyzak, C. Y. Suen, Methods of combining multiple classifiers and their
applications to handwriting recognition, IEEE Trans. System, Man, and Cybernetics,
27(3): 418-435, 1992.
13. J. Kittler, M. Hatef, R.P.W. Duin, J. Matas, On combining classifiers, IEEE Trans. Pat-
tern Analysis and Machine Intelligence, 20(3): 226-239, 1998.

19
14. J. Liu, Real Time Chinese Handwriting Recognition, E.E. Thesis, MIT, Cambridge, 1966.
15. M. Zobrak, A method for rapid recognition hand drawn line patterns, M.S. Thesis, Uni-
versity of Pittsburgh, 1966.
16. K. Yamamoto, A. Rosenfeld, Recognition of handprinted Kanji characters by a relaxation
method, Proc. 6th ICPR, Munich, 1982, pp.395-398.
17. K.S. Fu, Syntactic Methods in Pattern Recognition, Academic Press, 1974.
18. K.S. Fu, Syntactic Pattern Recognition and Applications, Prentice-Hall, 1982.
19. J.W. Tai, A syntactic-semantic approach for Chinese character recognition, Proc. 7th
ICPR, Montreal, Canada, 1984, pp.374-376.
20. F. Kimura, K. Takashina, S. Tsuruoka, Y. Miyake, Modified quadratic discriminant
functions and the application to Chinese character recognition, IEEE Trans. Pattern
Analysis and Machine Intelligence, 9(1): 149-153, 1987.
21. I.-J. Kim, J.H. Kim, Statistical character structure modeling and its application to hand-
written Chinese character recognition, IEEE Trans. Pattern Analysis and Machine Intelli-
gence, 25(11): 1422-1436, 2003.
22. J. Tsukumo, H. Tanaka, Classification of handprinted Chinese characters using non-linear
normalization and correlation methods, Proc. 9th ICPR, Rome, 1988, pp.168-171.
23. H. Yamada, K. Yamamoto, T. Saito, A nonlinear normalization method for hanprinted
Kanji character recognition--line density equalization, Pattern Recognition, 23(9): 1023-
1029, 1990.
24. Y.X. Gu, Q.R. Wang, C.Y. Suen, Application of a multilayer decision tree in computer
recognition of Chinese characters, IEEE Trans. Pattern Analysis and Machine Intelli-
gence, 5(1): 83-89, 1983.
25. Q.R. Wang, C.Y. Suen, Analysis and design of a decision tree based on entropy reduction
and its application to large character set recognition, IEEE Trans. Pattern Analysis and
Machine Intelligence, 6(4): 406-417, 1984.
26. J.W. Tai, Y.J. Liu, Chinese character recognition, Syntactic and Structural Pattern Rec-
ognition--Theory and Application, H. Bunke and A. Sanfeliu (Eds.), World Scientific,
1989.
27. H. Hao, X. Xiao, R. Dai, Handwritten Chinese character recognition by metasynthesis
approach, Pattern Recognition, 30(8), 1321-1328, 1997.
28. X. Lin, X. Ding, M. Chen, R. Zhang, Y. Wu, Adaptive confidence transform based clas-
sifier combination for Chinese character recognition, Pattern Recognition Letters, 19(10):
975-988, 1998.
29. K. Yamamoto, H. Yamada, T. Saito, R. Oka, Recognition of handprinted Chinese charac-
ters and Japanese cursive syllabury, Proc. 7th ICPR, Montreal, 1984, pp.385-388.
30. H. Yamada, Contour DP matching method and its application to handprinted Chinese
character recognition, Proc. 7th ICPR, Montreal, 1984, pp.389-392.
31. K. Yamamoto, H. Yamada, T. Saito, I. Sakaga, Recognition of handprinted characters in
the first level of JIS Chinese characters, Proc. 8th ICPR, Paris, 1986, pp.570-572.
32. J. Tsukumo, Handprinted Kanji character recognition based on flexible template match-
ing, Proc. 11th ICPR, The Hague, 1992, Vol.2, pp.483-486.
33. J. Guo, N. Sun, Y. Nemoto, M. Kimura, H. Echigo, R. Sato, Recognition of handwritten
characters using pattern transformation method with cosine function, Trans. IEICE Japan,
J76-D-II(4): 835-842, 1993 (in Japanese).
34. K. Saruta, N. Kato, M. Abe, Y. Nemoto, High accuracy recognition of ETL9B using
exclusive learning neural network-II (ELNET-II), IEICE Trans. Information and Systems,
79-D(5): 516-521, 1996.
35. M. Suzuki, S. Omachi, N. Kato, H. Aso, H. Nemoto, A discrimination method of similar
characters using compound Mahalanobis function, Trans. IEICE Japan, J80-D-II(10):
2752-2760, 1997 (in Japanese).

20
36. F. Kimura, T. Wakabayashi, S. Tsuruoka, Y. Miyake, Improvement of handwritten Japa-
nese character recognition using weighted direction code histogram, Pattern Recognition,
30(8): 1329-1337, 1997.
37. N. Kato, M. Suzuki, S. Omachi, H. Aso, Y. Nemoto, A handwritten character recognition
system using directional element feature and asymmetric Mahalanobis distance, IEEE
Trans. Pattern Analysis and Machine Intelligence, 21(3): 258-262, 1999.
38. K. Sawa, T. Wakabayashi, S. Tsuruoka, F. Kimura, Y. Miyake, Accuracy improvement
by gradient feature and variable absorbing covariance matrix in handwritten Chinese
character recognition, IEEE Trans. Pattern Analysis and Machine Intelligence, J84-D-
II(11): 2379-2397, 2001 (in Japanese).
39. J.X. Dong, A. Krzyzak, C.Y. Suen, High accuracy handwritten Chinese character recog-
nition using support vector machine, Proc. Int. Workshop on Artificial Neural Networks
for Pattern Recognition, Florence, Italy, 2003.
40. H. Liu, X. Ding, Handwritten character recognition using gradient feature and quadratic
classifier with multiple discrimination schemes, Proc. 8th ICDAR, Seoul, Korea, 2005,
pp.19-23.
41. C.-L. Liu, High accuracy handwritten Chinese character recognition using quadratic
classifiers with discriminative feature extraction, Proc. 18th ICPR, Hong Kong, 2006,
Vol.2, pp.942-945.
42. T. Horiuchi, R. Haruki, H. Yamada, K. Yamamoto, Two-dimensional extension of
nonlinear normalization method using line density for character recognition, Proc. 4th
ICDAR, Ulm, Germany, 1997, pp.511-514.
43. C.-L. Liu, K. Marukawa, Pseudo Two-dimensional shape normalization methods for
handwritten Chinese character recognition, Pattern Recognition, 38(12): 2242-2255, 2005.
44. A. Kawamura, K. Yura, T. Hayama, Y. Hidai, T. Minamikawa, A. Tanaka, S. Masuda,
On-line recognition of freely handwritten Japanese characters using directional feature
densities, Proc. 11th ICPR, The Hague, 1992, Vol.2, pp.183-186.
45. G. Srikantan, S.W. Lam, S.N. Srihari, Gradient-based contour encoder for character
recognition, Pattern Recognition, 29(7): 1147-1160, 1996.
46. C.-L. Liu, K. Nakashima, H. Sako, H. Fujisawa, Handwritten digit recognition: bench-
marking of state-of-the-art techniques, Pattern Recognition, 36(10): 2271-2285, 2003.
47. K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd edition, Academic
Press, 1990.
48. M. Loog, R.P.W. Duin, Linear dimensionality reduction via a heteroscedastic extension
of LDA: the Chernoff criterion, IEEE Trans. Pattern Analysis and Machine Intelligence,
26(6): 732-739, 2004.
49. C.-L. Liu, M. Nakagawa, Evaluation of prototype learning algorithms for nearest
neighbor classifier in application to handwritten character recognition, Pattern Recogni-
tion, 34(3): 601-615, 2001.
50. Y.Y. Tang, et al., Offline recognition of Chinese handwriting by multi-feature and multi-
level classification, IEEE Trans. Pattern Analysis and Machine Intelligence, 20(5): 556-
561, 1998.
51. Dai Ruwei, Hao Hongwei, Xiao Xuhong, Systems and Integration of Chinese Character
Recognition, Zhejiang Science and Technology Press, 1998 (in Chinese).
52. Dai Ruwei, Wang Lixin, Pattern recognition systems integration by metasynthesis, Sys-
tems Science and Systems Engineering, Scientific and Technical Documents House, Bei-
jing, 1997, 7-13.
53. B.H. Xiao, C.H. Wang, R.W. Dai, Adaptive combination of classifiers and its application
to handwritten Chinese character recognition, Proc. 15th ICPR, Barcelona, 2000, pp. 327-
330.

21
54. Wang Chunheng, Xiao Baihua, Dai Ruwei, Parallel compact integration in handwritten
Chinese character recognition, Science in China Series F--Information Sciences, 47(1):
89-96, 2004.
55. B.H. Xiao, C.H. Wang, R.W. Dai, Handwritten Chinese character recognition by meta-
synthetic approach, Int. J. Information Technology and Decision Making, World Scien-
tific Press, 1(4): 621-634, 2003.
56. K.S. Fu, Sequential Methods in Pattern Recognition and Machine Learning, Academic
Press, New York, 1968.
57. K.S. Fu, Pattern Recognition and Machine Learning, Plenum Press, 1971.
58. K.S. Fu, Grammatical inference: introduction and survey, Part I and Part II, IEEE Pattern
Analysis and Machine Intelligence, 8(3): 343-375, 1986.
59. K.C. You, K.S. Fu, A syntactic approach to stage recognition using attributed
grammars,IEEE Trans. System, Man, and Cybernetics, 9(6): 334-345, 1979.
60. W.-H. Tsai, K.S. Fu, Attributed grammar--a tool for combining syntactic and statistical
approaches to pattern recognition, IEEE Trans. System, Man, and Cybernetics, 10(12):
873-885, 1980.
61. W.-H. Tsai, K.S. Fu, A syntactic-statistical approach to recognition of industrial
objects,Proc. 5th ICPR, Miami, 1980, pp.251-259.
62. W.-H. Tsai, K.S. Fu, Error-correcting isomorphism of attributed relational graphs for
pattern analysis, IEEE Trans. System, Man, and Cybernetics, 9(12): 757-768, 1979.
63. W.-H. Tsai, K.S. Fu, A pattern deformation model and Bayes error-correcting recognition
system, IEEE Trans. System, Man, and Cybernetics, 9(12): 745-756, 1979.
64. J.W. Tai, K.S. Fu, Semantic syntax-directed translation for pictorial pattern recognition,
Technical Report, School of EE, Purdue University, TR-EE 81-83, Oct 1981.
65. D.E. Knuth, Semantics of context-free language, Journal of Mathematical System Theory,
2(2): 127-145, 1968.
66. J.W. Tai, A kind of relational attributed grammars, Acta Automatica Sinica, 9(2), 1983
(in Chinese).
67. Qian Xuesen, Yu Jingyuan, Dai Ruwei, A new discipline of science--the study of open
complex giant system and its methodology, Chinese J. System Engineering and Electron-
ics (in English), 4(2): 2-12, 1993.

22

View publication stats

You might also like