Neurocomputing 89 (2012) 213–219

Contents lists available at SciVerse ScienceDirect

Neurocomputing journal homepage: www.elsevier.com/locate/neucom

Letters

A new handwritten character segmentation method based on nonlinear clustering Jun Tan a, Jian-Huang Lai b,n, Chang-Dong Wang b, Wen-Xian Wang c, Xiao-Xiong Zuo c a

School of Mathematics and Computational Science, Sun Yat-sen University, Guangzhou, PR China School of Information Science and Technology, Sun Yat-sen University, Guangzhou, PR China c Public Security of Yuexiu, Guangzhou, PR China b

a r t i c l e i n f o

abstract

Article history: Received 3 November 2011 Received in revised form 13 January 2012 Accepted 15 February 2012 Communicated by Qi Li Available online 29 March 2012

In handwritten character recognition, it is a significant step to segment a text line into characters. The unsupervised clustering is a common approach for this task. However, due to the strong overlapping and touch among characters, the separation boundaries between two characters are usually nonlinear, which leads to the failure of the widely used clustering methods such as k-means. To tackle this problem, this paper proposes a new handwritten character segmentation method based on nonlinear clustering methods. In the proposed approach, we first segment the entire text line into strokes, the similarity matrix of which is computed according to stroke gravities. Then, the nonlinear clustering methods are performed on this similarity matrix to obtain cluster labels for these strokes. According to the obtained cluster labels, the strokes are combined to form characters. In this paper, we consider two nonlinear clustering methods, namely, spectral clustering based on Normalized cut (Ncut) and kernel clustering based on Conscience On-Line Learning (COLL). Whereby, two segmentation approaches are proposed with the one using Ncut termed SegNcut, and the one using COLL termed SegCOLL. Experiments on four databases are conducted to demonstrate the effectiveness of our SegNcut and SegCOLL approaches. & 2012 Elsevier B.V. All rights reserved.

Keywords: Handwritten character segmentation Overlapping Touch Nonlinearly separable Nonlinear clustering

1. Introduction Character recognition in freestyle handwritten documents has received a significant amount of attention in the past several decades [1–5]. It plays an important role in various fields such as forensic [1], historic document analysis [2], archaeology [4], etc. Given a freestyle handwritten document, two main steps of character recognition include segmentation and recognition, where segmentation is divided into two off-line components, namely, extracting text lines from page images and segmenting characters from text lines [6]. Segmentation is an indispensable step in character recognition. However, off-line segmentation is not an easy task for two reasons. First, the number of characters is quite huge (e.g., there are more than 9000 frequently used Chinese characters). Second, there exist various writing styles, which make both text line and single character segmentations very difficult. This paper mainly focuses on segmenting a text line comprising overlapping characters of arbitrary style into single characters.

n

Corresponding author. Tel.: þ86 13168313819. E-mail addresses: [email protected] (J. Tan), [email protected] (J.-H. Lai), [email protected] (C.-D. Wang). 0925-2312/$ - see front matter & 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.neucom.2012.02.026

Some methods have been proposed to segment characters recently. Tseng and Lee [7] presented a recognition-based character segmentation method for handwritten Chinese characters. This method firstly initializes several possible nonlinear segmentation paths by using a probabilistic Viterbi algorithm. These paths are further determined as candidate segmentation paths by verifying overlapping paths, between-character gaps, and adjacent-path distances. A segmentation graph is constructed using these candidate paths with the shortest path finally detected as the segmentation path. In [8], the strokes are used to build stroke bounding boxes (SBB), which are merged by some knowledgebased merging operations, and a dynamic programming approach is used to find the best segmentation boundaries. In [9], Yamaguchi et al. proposed a segmentation method for touching Japanese handwritten characters, which decreases oversegmentation by utilizing connecting condition of lines at the touching point. This method generates vertical straight lines to segment characters and is effective when characters are linearly separable, i.e., the characters can be partitioned via a straight line. However, in some handwriting documents of arbitrary style, the segmentation path between two characters cannot be separated by a straight line. Wang et al. [10] proposed a method for Chinese character segmentation based on connected components algorithm (CCA) template. This method first finds a black (foreground)

214

J. Tan et al. / Neurocomputing 89 (2012) 213–219

pixel as the center of 3  3 template, from which the search of neighbor foreground pixels begins, until no new neighbor foreground pixel can be found. Although it can well segment nonlinearly separable characters, its performance strongly relies on the definition of neighbor connected components and is sensitive to noise. In this paper, we propose a new character segmentation algorithm based on nonlinear clustering. The proposed approach first factorizes each text line into strokes, the similarity matrix of which is computed. Then the nonlinear clustering methods are applied on this similarity matrix to obtain cluster labels for these strokes. There exist many nonlinear clustering algorithms [11–14]. In this paper, we consider two state-of-the-art nonlinear clustering methods, namely, spectral clustering based on Normalized cut (Ncut) [11] and kernel clustering based on Conscience OnLine Learning (COLL) [12,13]. The segmented characters can be finally formed by combining strokes according to their cluster labels and positions. That is, the strokes with the same cluster label and similar positions make up one character. Since the clustering methods used are nonlinear, the resulting (implicit) segmentation paths cannot be straight, which feasibly adapts to the arbitrary writing styles. Additionally, no explicit separation path is generated, leading to the quite low computational complexity. The remainder of the paper is organized as follows. Section 2 introduces the preliminary knowledge of this work. In Section 3, we describe in detail the proposed method. Experimental results are reported in Section 4. We conclude our paper in Section 5.

2. Preliminary knowledge This section introduces the preliminary knowledge of this work, including the concepts involving character segmentation, the nonlinearly separable problem suffered by handwritten character segmentation and the nonlinear clustering methods such as spectral clustering [11] and kernel clustering [12,13]. Single character segmentation is an indispensable step of offline character recognition. Uncorrect segmentation will directly

lead to erroneous recognition. Due to the various writing styles, which often contains touch and overlapping between two characters, character segmentation is not an easy task. According to the distance and separability between two characters, character segmentation is classified into the following three categories.

 Type one: Characters are isolated from each other, as shown in  

Fig. 1(a). The radicals forming different characters can be easily separated according to which character they belong to. Type two: Characters do not touch with each other, but are not linearly separable by vertical straight lines due to overlapping between characters, as shown in Fig. 1(b). Type three: Characters touch with each other, as shown in Fig. 1(c).

Among the above three types, type two and type three are more challenging than type one. In this paper, we mainly focus on type two, i.e., the text lines contain characters that do not touch with each other, but are not linearly separable by vertical straight lines due to overlapping between them. The major difficulty of type two is that the separation gap between two characters are nonlinear, which cannot be achieved by a vertical straight line. Previous methods [8–10] are either on segmenting characters that are linearly separable using a vertical straight line or on directly finding nonlinear segmentation paths. As shown in Fig. 2, directly finding nonlinear segmentation paths is a NP hard task in that there exist plenty of candidate segmentation paths without mathematical model. One efficient remedy for selecting correct segmentation paths from candidates is clustering. The basic idea is to over-segment one text line into many strokes, forming a stroke set. Then a clustering method is performed on this stroke set to generate cluster labels for these strokes. Finally, these strokes are combined as characters according to their cluster labels. For instance, Wang and Fan [15] proposed to use the classic k-means algorithm [16] to cluster strokes. However, it can only deal with linearly separable characters due to the linear limitation of k-means. To this end, nonlinear clustering methods [11–14] should be used

Fig. 1. Three types of character segmentation. (a) Type one. (b) Type two. (c) Type three.

W

W1

X

char i

char i

H

H

char j

Overlap region

char j

A Curved Segmentation path

Candidate Segmentation path

Fig. 2. The overlap region and candidate segmentation paths. (a) Two overlapping characters and (b) candidate segmentation paths.

J. Tan et al. / Neurocomputing 89 (2012) 213–219

instead of linear clustering methods [16,17]. By using the eigenvectors corresponding to the largest eigenvalues to transfer the original stroke metric into the eigenvector space metric (spectral clustering) or using the nonlinear mapping f to transfer the original stroke metric into the high dimensional kernel space metric (kernel clustering), where the nonlinear cluster boundaries become linear, the two representative nonlinear clustering methods, i.e., spectral clustering [11] and kernel clustering [12], can efficiently identify the nonlinearly separable clusters. Thereby, these nonlinear clustering methods avoid the exhaustive search for the optimal nonlinear segmentation paths, which is NP hard as mentioned before.

two 3-fork points are merged into one 4-fork point, as shown in Fig. 5. By scanning the preprocessed image, all the strokes are extracted. Let fS1 ,S2 , . . . ,Sn g be n strokes obtained from an image C, where each stroke Si is a set of pixels. We first compute the gravity /Gkx ,Gky S of each stroke Sk !, ! M M X X Gkx ¼ i  P x ðiÞ Px ðiÞ i¼1

Gky ¼

N X i¼1

3. The proposed methods One method for tackling nonlinear separability in segmenting text lines into characters is to use nonlinear clustering methods. In this paper, we utilize two well-known nonlinear clustering methods, namely, spectral clustering based on Normalized cut (Ncut) [11] and kernel clustering based on Conscience On-Line Learning (COLL) [12,13], and respectively propose two character segmentation methods termed SegNcut and SegCOLL. Fig. 3 summarizes the overall flowchart of the proposed character segmentation methods, with some of the main procedures described as follows. The input of our approach is the image of characters, for instance, the scanned images of text, which are often gray-scale. We should first convert them into binary images with the background pixels (valued 0) being black and the foreground pixels (valued 1) being white. Then we utilize the thinning method proposed in [18] to obtain the skeleton of unitary thinness. After that, we obtain four types of feature points including 1-fork point (ending point), 2-fork point (intersecting point), 3-fork point (T-junction point), and 4-fork point (crossing point), as shown in Fig. 4. The redundant twigs are cut off and the corresponding Image of Character

Binary image of text

Skeleton &Spur

Merge Fork point

Compute gravity

Euclidean distance

Ncut/COLL

Extract Stroke

215

!, i  P y ðiÞ

N X

! P y ðiÞ

where M and N are the width and height of Sk respectively, and Px(i) and Py(j) are the foreground pixel number in the ith vertical and jth horizontal line, respectively. The Euclidean distance between two strokes Si and Sj are then defined as qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi JSi Sj J ¼ ðGix Gjx Þ2 þ ðGiy Gjy Þ2 ð3Þ After obtaining the Euclidean distance JSi Sj J which measures the spatial relation between Si and Sj, we perform two nonlinear clustering methods on the Euclidean distances of strokes to generate nonlinear cluster labels. That is, the boundary between two combined characters are not linear. The first nonlinear clustering method used is the spectral clustering based on Normalized cut (Ncut) [11]. To this end, we first transform the Euclidean distances of strokes into a similarity matrix W ¼ ½wij nn [11] ( 2 eðJSi Sj J Þ=sC if JSi Sj J o r wij ¼ ð4Þ 0 otherwise where sC is the variance of all the distances between strokes contained in image C and r is a threshold, say, the mean value of all distances. Then Ncut is performed as follows [11]: 1. First, an n  n diagonal matrix D is computed with its diagonal P element Dii ,i ¼ 1, . . . ,n being Dii ¼ nj¼ 1 wij . That is, the diagonal element Dii of D measures the connectivity from node i to all other nodes. 2. Obtain the matrix L ¼ D1=2 ðDWÞD1=2 .

Class label

Single characters

Fig. 3. Flowchart of the proposed character segmentation method.

ð2Þ

i¼1

Similarity matrix

Group by label

ð1Þ

i¼1

Fig. 5. Merging two 3-fork points to one 4-fork point.

Fig. 4. Illustration of four types of fork points. (a) Character, (b) 1-fork point, (c) 2-fork point, (d) 3-fork point, (e) 4-fork point.

216

J. Tan et al. / Neurocomputing 89 (2012) 213–219

3. Find the k (k being the number of clusters) eigenvectors of L, denoted as x1 ,x2 , . . . ,xk , corresponding to the k largest eigenvalues. From the matrix X ¼ ½x1 ,x2 , . . . ,xk  A Rnk by stacking the eigenvectors in columns, and normalize this matrix X to P have unit length by X ij ¼ X ij =ð nh ¼ 1 X 2ih Þ1=2 . 4. Take each row of X as an object in Rk and cluster these rows into k clusters via k-means. 5. Obtain the cluster label of each original stroke according to the cluster label of each corresponding row in X.

(b) Update the winning prototype W f ni ,: with learning rate Zt by 8 f > if j ¼ 1, . . . ,n, > > ð1Zt ÞW ni ,j þ Zt K i,j > < f 2 f 2 ð8Þ W ni ,j ’ ð1Zt Þ W ni ,j þ Zt K i,i > > > f > : þ 2ð1Zt ÞZt W n ,i if j ¼ n þ 1: i and the winning frequency ff k g by , c X nni ’nni þ1,f k ¼ nk nl , 8k ¼ 1, . . . ,c:

ð9Þ

l¼1

The second nonlinear clustering method used is the kernel clustering based on Conscience On-Line Learning (COLL) [12,13]. A kernel matrix K is firstly constructed by a Gaussian kernel which is the same as W computed as (4), i.e., K ¼W. Let n denote a cluster assignment mapping, i.e., n : X -f1, . . . ,cg. Then COLL is performed as follows [12,13]:

1. Randomly initialize assignment n and set the iteration index t ¼0. Initialize the winning frequency ff k g by nk ¼ 9n1 ðkÞ9,f k ¼ nk =n, 8k ¼ 1, . . . ,c, where nk denotes the winning times of the kth prototype. Initialize prototype descriptor W f A R f

W :,1:n ¼ AK,

f

W :,n þ 1 ¼ diagðAKA> Þ

cðn þ 1Þ

by ð5Þ

4. Compute ef via  c X ef ¼ 1 k¼1

þ Z2t

1 ð1Zt Þmk

mk X mk c X X

2

! f

W k,n þ 1

K pk , pk h

l

hþl k ¼ 1 h ¼ 1 l ¼ 1 ð1Zt Þ 0 1 f  X mk W c X k, pkl 1 @ 1 A: þ 2Zt ð1Zt Þmk l ¼ 1 ð1Zt Þl k¼1

ð10Þ

and return to step 2 until ef r E or t Zt max . 5. Obtain the cluster assignment ni , 8i ¼ 1, . . . ,n by

ni ¼ arg minfðK i,i þ W fk,n þ 1 2W fk,i Þg:

ð11Þ

k ¼ 1,...,c

where diagðMÞ denotes the main diagonal of a matrix M and the positive matrix A ¼ ½Ak,i cn has the form 8 1 > < if i A n1 ðkÞ ð6Þ Ak,i ¼ 9n1 ðkÞ9 > :0 otherwise 2. Get c empty index arrays fpk ¼ | : k ¼ 1, . . . ,cg and a random permutation fI1 , . . . ,In : Ii A f1, . . . ,ng s:t: Ii a Ij ,8i ajg, set t ¼ t þ 1. 3. For l ¼ 1, . . . ,n (a) Select the winning prototype W f ni ,: of the ith point (i ¼ Il ) by

ni ¼ arg minff k ðK i,i þ W fk,n þ 1 2W fk,i Þg: k ¼ 1,...,c

and append i to the ni th index array pni .

Therefore, we obtain the cluster labels fL1 ,L2 , . . . ,Ln g for the n strokes. These strokes are combined to form characters according to their cluster labels and positions, as shown in Fig. 6. For simplicity, the proposed character segmentation method with Ncut as the nonlinear clustering is termed SegNcut, and the proposed segmentation method with COLL as the nonlinear clustering is termed SegCOLL.

4. Experimental results

ð7Þ In this section, we conduct comparative experiments to validate the effectiveness of the proposed SegNcut and SegCOLL

Fig. 6. Combining strokes to form characters according to their cluster labels and positions.

J. Tan et al. / Neurocomputing 89 (2012) 213–219

methods on four widely tested databases. By comparing the correct rate and average segmentation time per character, the experimental results have shown the great improvement of our algorithms over existing methods.

4.1. Testing databases Four widely tested databases are used in our experiments. They are IAM [19], KAIST Hanja1 [20], HCL2000 [21] and HIT-MW [22]. The IAM database consists of handwritten English sentences based on the Lancaster-Oslo/Bergen (LOB) corpus. The database includes 1066 forms produced by 350 different writers and there are a total of 82 227 word instances out of a vocabulary of 10 841 words [19]. The KAIST Hanja1 database was collected by the Korea Advanced Institute of Science and Technology. It contains 783 frequently used Chinese characters, each character consisting of 200 samples written by 200 writers respectively [20]. The HCL2000 database was collected by Beijing University of Posts & Telecommunications and Pattern Recognition and Intelligent System Laboratory. The database contains 3755 frequently used simplified Chinese characters written by 1000 different writers [21]. The HIT-MW database was proposed by Harbin Institute of Technology. This database consists of 853 forms and 186 444 characters that are produced under an unconstrained condition

without preprinted character boxes [22]. The first two rows of Table 1 summarize the numbers of overall instances and writers respectively. Fig. 7 shows some samples of the four databases. Please note that, although the IAM database contains many characters requiring type three character segmentation, i.e., characters that touch with each other, we only use a subset consisting of strings that are composed of characters requiring type two character segmentation, i.e., characters that do not touch with each other, but are nonlinearly separable by vertical straight lines due to overlapping between characters. 4.2. Comparative results We compare the proposed method with five well-known methods including Viterbi algorithm (Viterbi) proposed in [7], Stroke Bounding Boxes method (SBB) proposed in [8], Connected Components Algorithm (CCA) proposed in [10], Project Profile Histogram (PPH) proposed in [9] and k-means proposed in [15]. Each of the compared methods is well-adjusted/trained to

Table 2 Comparing the correct rate (CR %) and average segmentation time (s) per character in seconds on four databases. Algorithms

Table 1 Summary of the testing databases. Databases

IAM

KAIST Hanja1

HCL2000

HIT-MW

#overall instances/characters #overall writers #strings tested #characters tested

82 227 350 156 964

156 600 200 233 1255

N/A 1000 301 1608

186 444 N/A 211 2711

217

Viterbi [7] SBB [8] CCA [10] PPH [9] k-means [15] SegNcut SegCOLL

IAM

KAIST Hanja1

HCL2000

HIT-MW

CR

Time

CR

Time

CR

Time

CR

Time

79.2 82.7 75.4 73.5 61.2 85.4 88.1

1.351 1.271 1.244 1.115 1.746 1.102 0.983

77.4 81.3 74.3 72.6 46.2 89.2 85.2

1.353 1.221 1.644 1.334 1.819 1.162 1.216

79.2 81.3 77.1 70.8 57.2 88.9 89.3

1.353 1.241 1.544 1.417 1.525 1.189 1.175

77.8 82.9 75.3 72.6 51.2 89.7 87.7

1.458 1.624 1.332 1.749 1.847 1.025 1.144

Fig. 7. Some samples of the four databases. (a) IAM [19], (b) KAIST Hanja1 [20], (c) HCL2000 [21], (d) HIT-MW [22].

218

J. Tan et al. / Neurocomputing 89 (2012) 213–219

2

Correct rate (%)

80 60 Viterbi SBB CCA PPH k−means SegNcut SegCOLL

40 20 0

English

Chinese

Average time per character (sec.)

100

1.5

1

Viterbi SBB CCA PPH k−means SegNcut SegCOLL

0.5

0

English

Character type

Chinese Character type

Fig. 8. Overall comparison in segmenting both English and Chinese characters. (a) Correct rate and (b) average time per character.

generate the best results. Both the correct rate and average segmentation time per character are reported and compared. On each database, we select only text lines (strings) consisting of nonlinearly separable characters that are hard to segmented by vertical straight lines. On the IAM database, we obtain 156 strings comprising 964 characters. On the KAIST Hanja1 database, we obtain 233 strings consisting of 1255 characters. On the HCL2000 database, we obtain 301 strings consisting of 1608 characters. On the HIT-MW database, we obtain 211 strings consisting of 2711 characters. The last two rows of Table 1 summarize the numbers of tested strings and characters in each database, respectively. The comparative results on four databases are listed in Table 2. From the table, we observe that the proposed segmentation approaches, namely, SegNcut and SegCOLL, always produce the best segmentation results. On segmenting English characters, i.e., IAM dataset, the SegCOLL method is slightly better than SegNcut. On the remaining three Chinese character datasets, i.e., KAIST Hanja1, HCL2000 and HIT-MW, the SegNcut performs slightly better than SegCOLL. Compared with the existing segmentation, i.e., Viterbi, SBB, CCA, PPH and k-means, both of our approaches obtain a very significant improvement, especially on segmenting nonlinearly separable Chinese datasets, e.g., HIT-MW. Additionally, by comparing the computational time used to segment these characters, the proposed nonlinear approaches still have great advantages over the existing methods. As a overall comparison, Fig. 8 plots the comparative results on all strings collected from four databases according to the character types (English and Chinese), respectively. By comparison, our approaches are effective in segmenting both English and Chinese characters. The experimental results have demonstrated the effectiveness and efficiency of our approaches.

5. Conclusions In this paper, we present a new handwritten character segmentation method, which can effectively segment nonlinearly separable characters. In the proposed approach, we first segment the entire text line into strokes, and compute their similarity matrix according to the stroke gravities. Then, the nonlinear Normalized cut (Ncut) and Conscience On-Line Learning (COLL) clustering methods are performed, respectively, on this similarity matrix to obtain cluster labels for these strokes, through which the strokes are combined to form characters. Whereby, two segmentation approaches are proposed with the one using Ncut termed SegNcut, and the one using COLL termed SegCOLL.

Comparative results with some well-known methods on four testing databases demonstrate the effectiveness and efficiency of our SegNcut and SegCOLL approaches.

Acknowledgments This work was supported by the NSFC (61173084), the Technology Program of Guangdong (2009B060700091) and the Technology Program of Yuexiu district of Guangzhou. References [1] M. Bulacu, L. Schomaker, Text-independent writer identification and verification using textural and allographic features, IEEE Trans. Pattern Anal. Mach. Intell. 29 (2007) 701–717. [2] G. Zhu, Y. Zheng, D. Doermann, S. Jaeger, Signature detection and matching for document image retrieval, IEEE Trans. Pattern Anal. Mach. Intell. 31 (2009) 2015–2031. [3] Y. Liu, Z. You, L. Cao, A novel and quick SVM-based multi-class classifier, Pattern Recognit. 39 (2006) 2258–2264. [4] M. Panagopoulos, C. Papaodysseus, P. Rousopoulos, D. Dafi, S. Tracy, Automatic writer identification of ancient Greek inscriptions, IEEE Trans. Pattern Anal. Mach. Intell. 31 (2009) 1404–1414. [5] Y. Liu, S.S. Ge, C. Li, Z. You, k-NS: a classifier by the distance to the nearest subspace, IEEE Trans. Neural Networks 22 (2011) 1256–1268. [6] G. Louloudis, B. Gatos, I. Pratikakis, C. Halatsis, Text line and word segmentation of handwritten documents, Pattern Recognit. 42 (2009) 3169–3183. [7] Y.-H. Tseng, H.-J. Lee, Recognition-based handwritten Chinese character segmentation using a probabilistic Viterbi algorithm, Pattern Recognit. Lett. 20 (1999) 791–806. [8] L.Y. Tseng, R.C. Chen, Segmenting handwritten Chinese characters based on heuristic merging of stroke bounding boxes and dynamic programming, Pattern Recognit. Lett. 19 (1998) 963–973. [9] T. Yamaguchi, T. Yoshikawa, T. Shinogi, S. Tsuruoka, M. Teramoto, A segmentation method for touching Japanese handwritten characters based on connecting condition of lines, in: Proceedings of the Sixth International Conference on Document Analysis and Recognition, pp. 837–841. [10] L.-W. Wang, Y. Yang, B. Xie, Y. Yang, Segmentation of handwritten Chinese characters based on connected components algorithm and penetration times algorithm, Inf. Technol. 28 (2004) 31–35. [11] J. Shi, J. Malik, Normalized cuts and image segmentation, IEEE Trans. Pattern Anal. Mach. Intell. 22 (2000) 888–905. [12] C.-D. Wang, J.-H. Lai, J.-Y. Zhu, A conscience on-line learning approach for kernel-based clustering, in: Proceedings of the 10th International Conference on Data Mining, pp. 531–540. [13] C.-D. Wang, J.-H. Lai, J.-Y. Zhu, Conscience online learning: an efficient approach for robust kernel-based clustering, Knowl. Inf. Syst. 31 (1) (2012) 79–104. [14] C.-D. Wang, J.-H. Lai, J.-Y. Zhu, Graph-based multiprototype competitive learning and its applications, IEEE Trans. Syst. Man Cybern., Part C: Appl. Rev., doi:http://dx.doi.org/10.1109/TSMCC.2011.2174633. [15] A.-B. Wang, K.-C. Fan, Optical recognition of handwritten Chinese characters by hierarchical radical matching method, Pattern Recognit. 34 (2001) 15–35. [16] J. MacQueen, Some methods for classification and analysis of multivariate observations, in: Proceedings of the Fifth Berkeley Symposium on

J. Tan et al. / Neurocomputing 89 (2012) 213–219

[17] [18] [19]

[20]

[21]

[22]

Mathematical Statistics and Probability, vol. 1, University of California Press, 1967, pp. 281–297. C.-D. Wang, J.-H. Lai, Energy based competitive learning, Neurocomputing 74 (2011) 2265–2275. T.Y. Zhang, C.Y. Suen, A fast parallel algorithm for thinning digital patterns, Commun. ACM 27 (1984) 236–239. U.-V. Marti, H. Bunke, The IAM-database: an English sentence database for offline handwriting recognition, Int. J. Document Anal. Recognit. 5 (2002) 39–46. I.-J. Kim, J.-H. Kim, Statistical character structure modeling and its application to handwritten Chinese character recognition, IEEE Trans. Pattern Anal. Mach. Intell. 25 (2003) 1422–1436. H. Zhang, J. Guo, G. Chen, C. Li, HCL2000—a large-scale handwritten Chinese character database for handwritten character recognition, in: Proceedings of the 10th International Conference on Document Analysis and Recognition. T. Su, T. Zhang, D. Guan, Corpus-based HIT-MW database for offline recognition of general-purpose Chinese handwritten text, Int. J. Document Anal. Recognit. 10 (2007) 27–38.

Jun Tan received the BSc and MSc degrees in computational mathematics from Sun Yat-sen University, Guangzhou, PR China, in 1995 and 2001, respectively. He started the pursuit of the PhD degree with the School of Information Science and Technology of Sun Yat-Sen University in September 2009. He did teaching and research in the department of scientific computation and computer application, Sun Yat-sen University, Guangzhou, PR China, from 2002. Tan is currently working on developing statistical pattern recognition methods for automatic writer identification and for handwritten historical document retrieval. His scientific interests include computer vision, statistical pattern recognition, biometrics, and document analysis and recognition.

Jian-Huang Lai received his MSc degree in applied mathematics in 1989 and his PhD in mathematics in 1999 from Sun Yat-sen University, China. He joined Sun Yat-sen University in 1989 as an Assistant Professor, where currently, he is a professor with the Department of Automation of School of Information Science and Technology and vice dean of School of Information Science and Technology. His current research interests are in the areas of digital image processing, pattern recognition, multimedia communication, wavelet and its applications. He has published over 80 scientific papers in the international journals and conferences on image processing and pattern recognition, e.g. IEEE TNN, IEEE TIP, IEEE TSMC (Part B), Pattern Recognition, ICCV, CVPR and ICDM. Lai serves as a standing member of the Image and Graphics Association of China and also serves as a standing director of the Image and Graphics Association of Guangdong.

219

Chang-Dong Wang received the BS degree in applied mathematics in 2008 and MSc degree in computer science in 2010 from Sun Yat-sen University, China. He started the pursuit of the PhD degree with Sun Yat-sen University in September 2010. He is a visiting student at University of Illinois at Chicago with supervisor Philip S. Yu, from January 2012 to January 2013. His current research interests include machine learning and data mining, especially focusing on data clustering and its applications. He has published several scientific papers in international journals and conferences such as Neurocomputing, KAIS, IEEE TSMC-C, IEEE TKDE, and ICDM. His ICDM 2010 paper won the Honorable Mention for Best Research Paper Awards. He won the Student Travel Award from ICDM 2010 and ICDM 2011.

Wen-Xian Wang received his B.Sc. degree in Criminal Technology Department in 1999 from China Criminal Police University, P.R. China. Mr. Wang is currently working in Public Security of Yuexiu, Guangzhou, PR China. He is a Senior Engineer of document inspection. Mr. Wang current research interests include handwriting examination, document forming examination, destroyed document examination, language examination and identification, printed document examination.

Xiao-Xiong Zuo received his B.Sc. degree in Criminal Technology Department in 2003 from China Criminal Police University, P.R. China. Mr. Zuo is currently working in Public Security of Yuexiu, Guangzhou, PR China . He is an Engineer of document inspection. Mr. Zuo current research interests include handwriting examination, document forming examination, printed document examination.