Recent Advances in Handwriting Recognition · Recent Advances in Handwriting Recognition 3 • The...

Document Analysis, Editors: Umapada Pal, Swapan K. Parui, Bidyut B. Chaudhuri, pp.1-30

1

Recent Advances in Handwriting Recognition Flávio Bortolozzi, Alceu de Souza Britto Jr., Luiz S. Oliveira and Marisa Morita

Pontifical Catholic University of Paraná (PUCPR) Rua Imaculada Conceição, 1155 – CEP 80215-901

Curitiba - Paraná - Brazil

Abstract

Machine simulation of human reading has been subject of intensive research for the last three decades. This paper presents a summary about the recent advances in terms of character, word, numeral string, and setence recognition. In addition, the main new trends in the field of handwriting recognition are discussed and some important contributions are presented.

INTRODUCTION

In spite of the major effort that has been expended to bring about a paper-free society, a very large number of paper-based documents are processed daily by computers all over the world in order to handle, retrieve, and store information. The problem is that the manual process used to enter the data from these documents into computers demands a great deal of time and money. The field of Document Analysis and Recognition (DAR) has played a very important role in the attempt to overcome this problem. The general objective of DAR research is to fully automate the process of entering and understanding printed or handwritten data into the computer. Machine simulation of human reading has been the subject of intensive research for the last three decades. However, the early investigations were limited by the memory and power of the computer available at that time. With the explosion of information technology, there has been a dramatic increase of research in this field since the beginning of 1980s. The interest devoted to this field is not explained only by the exciting challenges involved, but also the huge benefits that a system, designed in the context of a commercial application, could bring.

Document Analysis

2

According to the way handwriting data is generated, two different approaches can be distinguished: on-line and off-line. In the former, the data are captured during the writing process by a special pen on an electronic surface. In the latter, the data are acquired by a scanner after the writing process is over. In this case, the recognition of off-line handwriting is more complex than the on-line case due to the presence of noise in the image acquisition process and the loss of temporal information such as the writing sequence and the velocity. This information is very helpful in a recognition process. Off-line and on-line recognition systems are also discriminated by the applications they are devoted to. The off-line recognition is dedicated to bank check processing, mail sorting, reading of commercial forms, etc, while the on-line recognition is mainly dedicated to pen computing industry and security domains such as signature verification and author authentication. In this work, we are concerned with the off-line recognition. This paper presents a summary about the recent advances in terms of character, word, numeral string, and setence recognition. In addition, the main new trends in the field of handwriting recognition are discussed and some important contributions are presented. For this purpose, this paper is organized as follows: Section 2 presents a summary of the recognition strategies used in the DAR research field. Sections 3 through 6 are dedicated to present the main contributions on character, numeral string, word and setence recognition. Section 7 presents the main investigations towards performance: feature selection, combination of classifiers and the use of sinthetical data. Finally, Section 8 concludes this study. RECOGNITION STRATEGIES

Numerous techniques for off-line handwriting recognition have been investigated based on four general approaches of pattern recognition, as suggested by Jain (Jain, 2000): template matching, statistical techniques, structural techniques, and neural networks. Such approaches are neither necessary independent nor disjointed from each other. Occasionally, a technique in one approach can also be considered to be a member of other approaches. Template matching operations determine the degree of similarity between two vectors (groups of pixels, shapes, curvatures, etc) in the feature space. Matching techniques can be grouped into three classes: direct matching (Gader, 1991), deformable templates and elastic matching (Dimauro, 1997), and relaxation matching (Xie, 1988). Statistical techniques are concerned with statistical decision functions and a set of optimal criteria, which determine the probability of the observed pattern belonging to a certain class. Several popular handwriting recognition approaches belong to this domain:

Recent Advances in Handwriting Recognition

3

• The k-Nearest-Neighbor (k-NN) rule is a popular non-parametric recognition method, where a posteriori probability is estimated from the frequency of nearest neighbors of the unknown pattern. Compelling recognition results for handwriting recognition have been reported using this approach (Guillevic, 1995). The drawback of this method is the high computational cost when the classification is conducted. To surpass such a problem some researchers have proposed faster k-NN methods. A comparison of fast nearest neighbor classifiers for handwriting recognition is given in (Mico, 1999).

• The Bayesian classifier assigns a pattern to a class with the maximum a posteriori probability. Class prototypes are used in the training stage to estimate the class-conditional probability density function for a feature vector (Duda, 2001).

• The polynomial discriminant classifier assigns a pattern to a class with the maximum discriminant value which is computed by a polynomial in the components of a feature vector. The class models are implicitly represented by the coeficients in the polynomial (Schurmann, 1996).

• Hidden Markov Model (HMM) is a doubly stochastic process, with an underlying stochastic process that is not observable (hence the word hidden), but can be observed through another stochastic process that produces the sequence of observations (Rabiner, 1989). An HMM is called discrete if the observations are naturally discrete or quantized vectors from a codebook or continuous if these observations are continuous. HMMs have been proven to be one of the most powerful tools for modeling speech and later on a wide variety of other real-world signals. These probabilistic models offer many desirable properties for modeling characters or words. One of the most important properties is the existence of eficient algorithms to automatically train the models without any need of labeling presegmented data. HMMs have been extensively applied to handwritten word recognition (El Yacoubi, 1999) (Kundu, 2002) (Senior, 2002) (Grandidier, 2003) and their applications to handwritten digit recognition (Cai, 1999) (Britto, 2001) have been growing. The literature presents two basic approaches for handwriting recognition using HMM: Model-Discriminant HMM and Path-Discriminant HMM. In the former, a model is constructed for each class (word, character, or segmentation unit) in the training phase. In the latter, a single HMM is constructed for the whole language or context. The performances of these two approaches are compared in various experiments by utilizing different lexicon sizes (Kundu, 2002).

• Fuzzy set reasoning is a technique that employs fuzzy set elements to describe the similarities between the features of the characters. Fuzzy set elements give more realistic results when there is not a priori knowledge about the data, and therefore, the probabilities cannot be calculated. The literature reports different approaches based on this technique such as fuzzy graphs (Abuhaiba, 1993), fuzzy rules (Gader, 1995), and linguistic fuzzy (Lazzerini, 2000).

Document Analysis

4

• Support Vector Machine (SVM) is based on the statistical learning theory (Vapnik, 1995) and quadratic programming optimization. An SVM is basically a binary classifier and multiple SVMs can be combined to form a system for multi-class classification. In the past few years, SVM has received increasing attention in the community of machine learning due to its excellent generalization performance. More recently, some SVM classification systems have been developed for handwriting digit recognition, and some promising results have been reported in (Ayat, 2002), (Byun, 2002) and (Oliveira 2004).

In structural techniques the characters are represented as unions of structural primitives. It is assumed that the character primitives extracted from handwriting are quantifiable, and one can find the relationship among them. Basically, structural methods can be categorized into two classes: grammatical methods (Shridhar, 1986) and graphical methods (Kim, 1998). A Neural Network (NN) is defined as a computing structure consisting of a massively parallel interconnection of adaptative “neural” processors. The main advantages of neural networks lies in the ability to be trained automatically from examples, good performance with noisy data, possible parallel implementation, and eficient tools for learning large databases. NNs have been widely used in this field and promising results have been achieved, especially in handwriting digit recognition. The most widely studied and used neural network is the Multi-Layer Perceptron (MLP) (Bishop, 1995). Such an architecture trained with back-propagation (LeCun, 1998a) is among the most popular and versatile forms of neural network classifiers and is also among the most frequently used traditional classifiers for handwriting recognition. See (Zhang, 2000) for a review. Other architectures include Convolutional Network (CN) (LeCun, 1998b), Self-Organized Maps (SOM) (Zhang, 1999), Radial Basis Function (RBF) (Bishop, 1995), Space Displacement Neural Network (SDNN) (Matan, 1992), Time Delay Neural Network (TDNN) (Lethelier, 1995), Quantum Neural Network (QNN) (Zhou, 1999), and Hopfield Neural Network (HNN) (Ling, 1997). The above review indicates that there are many recognition techniques available for handwriting recognition systems. All of them have their own advantages and drawbacks. In the recent years, many researchers have combined such techniques in order to improve the recognition results. The idea is not rely on a single decision making scheme. Various classifier combination schemes have been devised and it has been experimentally demonstrated that some of them consistently outperform a single best classifier (Kittler, 1998)( Xu, 2002). Another strategy that can increase the recognition rate in a relatively easy way with a small additional cost is through the use of verification. Such a scheme consists of refining the top few candidates in order to enhance the recognition rate economically. Such a kind of scheme


5

has been successfully applied to handwriting recognition in (Zhou, 1999)(Britto, 2001)(Koerich, 2002)(Oliveira, 2002). In summary, the current stage in the evolution of handwriting processing results from a combination of several elements, such as: the use of different strategies for feature selection, the evaluation of different classifiers and their combination, the use of complex systems based on contextual information, verifiers and postprocessing, the use of synthetic data, the automatic optimization of entire systems, and so forth. CHARACTER RECOGNITION The recognition of digits, a subfield of character recognition, is subject of much attention since the first years of research in the field of handwriting recognition. The motivation is the potential number of applications, and also because the solutions for this subset of characters are simpler and robust. Even nowadays, many approaches to solve this problem have been proposed in the literature. The positive aspect is that most of the knowledge provided by these investigations may be applied to character, word, and sentence recognition. Drawing up a taxonomy of the approaches for digit recognition available in the literature is difficult, since their methodologies overlap. However, research in this field has basically considered: a) feature extraction methods, b) classification methods, and c) system architectures based on different strategies, such as combinations of multiple classifiers, the use of multiple templates, and the use of verification modules. The investigation of feature extraction methods has gained considerable attention since a discriminative feature set is considered the most important factor in achieving high recognition performance. Trier et al (Trier, 1996) present an interesting survey of feature extraction methods for off-line recognition of segmented characters. The authors describe important aspects that must be considered before selecting a specific feature extraction method. In the same direction, an interesting review of shape analysis techniques is presented by Loncaric in (Loncaric, 1998). In general, the feature extraction methods for digit recognition reported in the literature are based on two types of features: statistical and structural. The statistical features are derived from statistical distributions of points, such as zoning, moments, projection histograms, or direction histograms (Kimura, 1992) (Gader, 1996a) (Cheung, 1998). Structural features are based on topological and geometrical properties of the character, like strokes and their directions, end-points, or intersections of segments, and loops (Hirano, 1997) (Cai, 1998). Many researchers have explored the integration of structural and statistical information to highlight different character properties, since these types of features are considered to be complementary (Cai, 1998) (Heute, 1998). In a recent work, Britto et al (Britto, 2004)

Document Analysis

6

combine foreground and background information extracted from columns and rows of character images (see Figure 1) to contemplate isolated digit and numeral string recognition. These complementary features are combined in a two-stage HMM-based method. Alongside these investigations of feature extraction methods, many other studies have address classification methods. Different classifiers have been used for handwritten digit recognition, such as statistical (Cheung, 1998) (Park, 1998), structural (Hirano, 1997) and neural networks (Zhang, 1998). Recently, significant contributions towards the improvement of recognition rates have been made by means of different combination strategies (Suen, 1999a) (Liu, 2004a), and the use of Support Vector Machines (Teow, 2002) (Milgram, 2004).

22222XXXXXXX1111111111111112222XXXXXXXXXX1111111111111222XXXXXXXXXXXXX11111111111XXXXXXXXXXXXXXXXX1111111111XXXXXXXXXXXXXXXXXX111111111XXXXXXXXXXXXXXXXXXX11111111XXXXXDDDDDDDDXXXXXXX11111113XXXDDDDDDDDDDDXXXXXX111111333777777777777XXXXXX111111333777777777777XXXXXX11111133377777777777XXXXXXX11111133377777777777XXXXXXX11111133377777777777XXXXXXX1111113337777777777XXXXXXXX111111333777777777XXXXXXXXX111111333777777777XXXXXXXXXX11111333777777XXXXXXXXXXXXX11111333777777XXXXXXXXXXXXXX111133377777XXXXXXXXXXXXXXX111133377777XXXXXXXXDDDXXXXX11133377777XXXXXXDDDDDDXXXX111333777777XXXXDDDDDDDDXXXX11333777777777777777777XXXX11333777777777777777777XXXXX1333777777777777777777XXXXX1333777777777777777777XXXXXX333777777777777777777XXXXXX333777777777777777777XXXXXX3337777777777777777XXXXXXX433377777777777XXXXXXXXXXXX4333XXXXXXXXXXXXXXXXXXXXXXX4333XXXXXXXXXXXXXXXXXXXXXX44333XXXXXXXXXXXXXXXXXXXXX4443333XXXXXXXXXXXXXX4444444443333XXXXXX444444444444444443333333XX444444444444444444

s 1s 2

s 3s 4

s 5ss m

ss1 s2 s3 sn

Fig.1 Complementary information extracted from columns and rows of character images

and combined in an HMM-based method (Britto, 2001)

Besides, investigations of new strategies for handwritten digit recognition have led researchers to figure out alternative ways to treat the variability intrinsic to handwritten numerals. In this context, three examples are interesting: the set of 23 digit prototypes and a deformable model-based recognition in a Bayesian framework proposed in (Cheung, 1998); the rule-based structural classifier used as a verification module (VM) for a Multilayer Feedforward Locally Connected Network presented in (Zhou, 1997); and, 3) the flexible image matching with 10,000 templates proposed in (Belongie, 2002).


7

Table 1 Some important contributions in the field of isolated digit recognition

Method RR (%)

Error (%)

Rejection (%)

Database Strategy

(Ha, 1997) 99.09 0.91 0 CEDAR (2,213)

Perturbation method

(Cai, 1998) 96.16 3.84 0 CEDAR (2,711)

Statistical and structural features combined in an HMM-based method

(Delevski, 1998)

99.37 0.27 0.36 CEDAR (2,213)

Morphological and Topological properties (graph representation)

(Suen, 1999a) 98.85 99.77

2.13 0.23

0 0

CENPARMI (2,000) CEDAR (2,213)

Multiples features and multiple neural networks

(Liu, 1999) 98.45 1.55 0 CENPARMI (2,000)

Neural networks and discriminative training

(Dong, 2001a)

99.01 0.99 0 MNIST (10,000)

Multi-net learning framework

(Teow, 2002) 99.41 0.59 0 MNIST (10,000)

Trio-wise linear SVM with direction and stroke features

(Liu, 2002) 99.58 0.42 0 MNIST (10,000)

SVM

(Belongie, 2002)

99.37 0.63 0 MNIST (10,000)

Flexible image matching with 10,000 templates

(DeCoste, 2002)

99.58 0,42 0 MNIST (10,000)

SVM

(Oliveira, 2004)

99.20 0.80 0 NIST SD19 (60,089)

SVM

(Britto, 2004) 98.00 2.00 0 NIST SD19 (60,089)

Combination of complementary features in an HMM-based method

A recent bechmarking of state-of-the-art techniques for digit recognition is presented by Liu et al in (Liu, 2003). The authors have used well-known databases (CENPARMI, CEDAR, and, MNIST) to evaluate different feature sets and classification techniques, which were proposed in the last years for isolated digit recognition. The experiments have shown that chaincode feature and the gradiente feature show advantage over other features, and the profile structure feature shows efficiency as a complementary feature. They also conclude that SVMs using Gaussian kernel outperform all traditional techniques such as neural networks (MLP and RBF), polynomial classifiers, and learning quadratic discriminant functions. Nevertheless, they point out that memory space and computational speed for classification still are important issues to be considered when discussing SVMs. Thus, some authors have proposed using SVMs for verification rather than classification. This strategy is computationally cheaper once SVMs are called just to solve difficult cases, which were

Document Analysis

8

rejected during the classification process. Table 1 summarizes some important contributions in the field of isolated digit recognition. It is important to point out that the recognition, rejection, and substitution rates reported are not directly comparable, since these results are based on different databases. As one can see, some maturity can be observed for isolated digit recognition. However, when we talk about the recognition of alphabetic characters, the problem becomes more complicated. The most obvious difference is the number of classes that can be up to 52, depending if uppercase (A–Z) and lowercase (a–z) characters are distinguished from each other. Consequently, there is a larger number of ambiguous alphabetic characters other than numerals. Character recognition is further complicated by other differences such as multiple patterns to represent a single character, cursive representation of letters, and the number of disconnected and multi–stroke characters (Koerich, 2002). In fact, we can say that character recognition still an open problem. Table 2 presents some recent contributions to the filed of character recognition evaluated on NIST database.

Table 2 Some recent contributions to the filed of character recognition

Method Training set Validion set Test set RR (%)

(Oh, 1998) Uppercase (26 classes)

26,000 - 11,941 90.0

(Dong, 2001b) Lowercase (26 classes)

23,937 - 10,688 92.3

(Koerich, 2002) Uppercase (26 classes) 37,440 12,092 11,941 92.3 Lowercase (26 classes) 37,440 11,578 12,000 84.6 upper/lower (52 classes) 74,880 23,670 23,941 85.5 (Britto, 2004) Uppercase (26 classes) 37,440 12,092 11,941 90.0 Lowercase (26 classes) 37,440 11,578 12,000 84.0 upper/lower (52 classes) 74,880 23,670 23,941 87.0

Fig.2 Difficulties in recognizing numeral strings

Broken numerals

Overlapping numerals

Touching numerals

Slanted numerals

Size variation


9

NUMERAL STRING RECOGNITION An important subject of research in the field of document analysis and recognition has been the recognition of numeral strings. The principal motivation is the wide variety of potential applications, such as, ZIP codes, bank checks, tax forms, and census forms. The challenge is to recognize numeral strings of unknown length which are not neatly written. Some possible difficulties contributing to the unsatisfactory performance of many methods for recognizing handwritten numeral strings are shown in Figure 2. One of the main bottlenecks in a handwriting numeral string recognition system is the segmentation step, which can be found in different levels, e.g., segment the information of interest from the document, segment phrases into words, and words into characters. Let us focus in the latter. Consider for example the problem of handwritten digit string recognition. In that case, the segmentation consists in reading a string of digits (sometimes non-digits can appear, e.g., systems to read bank cheques) and then to segment this string into pieces. The main problem lies in the fact that usually we do not know the number of digits in the string, hence, the optimal boundary between two digits is unknown.

C0

C1

C2

C3

SP0

SP1

SP2

Fig.3 Example of oversegmentation

During the last 15 years, several authors have devoted efforts in order to build more robust segmentation algorithms, which can be divided into two classes: segmentation-then-recognition (sometimes called segmentation-free) and segmentation-based recognition. In the former, the segmentation occurs before recognition while in the latter each segmentation

Document Analysis

10

hypothesis is evaluated by the recognition module. This makes the segmentation-based recognition a more robust strategy, but more time consuming as well. Segmentation-based recognition can be explicit when based on cut rules or implicit when each pixel column is a potential cut location. In the case of explicit segmentation several algorithms have been proposed during the last years. They normally take into consideration a set of heuristics and information of the foreground (Lethelier, 1995)(Yu, 2001)(Kim, 2002), background (Lu, 1999)(Chen, 2000)(Pal, 2003), or a combination of them (Oliveira, 2002) in order to generate potential segmentation cuts. Figure 3 shows an example of oversegmentation provided by the method described in (Oliveira, 2002). The main problems of these algorithms are the elevated number of potential cuts, which must be evaluated by the recognition algorithm, and the number of heuristics that must be set. In addition, most of the time a segmentation-based method do not represent well a good compromise between segmentation and recognition processes. In fact, a correct recognition often depends on a correct segmentation. By contrast, a correct segmentation also requires a correct recognition. Thus, we can say that these statements have a “chicken and egg” relationship, and they should be approached simultaneously. An alternative aimed at avoiding the prior segmentation of the string has been the use of implicit segmentation-based methods to integrate segmentation and recognition processes. A promising approach to achieve this has been based on Hidden Markov Models (HMMs). This approach was originally developed in the field of speech recognition (Rabiner, 1989), where it has been applied with much success. The benefits of applying such a technique to recognize printed words have been shown in (Bose, 1994) (Elms, 1998). In (Procter, 1998), the method proposed in (Elms, 1998) was adapted for handwritten numeral strings. From these studies, we may conclude that such an approach is a promising way of integrating segmentation and recognition to deal with the difficulties encountered in processing handwritten numeral strings. However, we can also observe some cost attached to this integration, which is an open problem in such an approach. This cost is a loss in recognition performance caused by combining segmentation with recognition. In other words, the problem is that a set of features and models that show promising performance in terms of segmentation usually do not show similar performance in terms of recognition, and vice-versa. By contrast, to integrate them it is necessary to define features and numeral models to contemplate both the segmentation and recognition aspects simultaneously. Moreover, a feature set must be extracted from a numeral string image in the same way that it is from an isolated digit image. In summary, the challenge is to find some way to compensate for the loss in recognition performance resulting from the necessary tradeoff between segmentation and recognition carried out in an implicit segmentation-based method. In a recent work, Britto et al (Britto, 2003) propose a way of combining segmentation and recognition taking into account this necessary tradeoff. The method is based on a two-stage recognition strategy


11

Table 3 Performance of numeral string recognition based data in NIST SD19

Method String class RR (%) Error (%)

Rejection (%)

# of samples on the testing set

2_digit 94.20 1.00 4.80 1,000 3_digit 87.90 1.00 11.10 1,000 4_digit 79.90 1.00 19.10 1,000 5_digit 75.60 1.00 23.40 1,000 6_digit 63.30 1.00 35.70 1,000

(Matin, 1993)

10_digit NA NA NA NA 2_digit 96.20 3.80 0.00 981 3_digit 92.70 7.30 0.00 986 4_digit 93.20 6.80 0.00 988 5_digit 91.10 8.90 0.00 988 6_digit 90.30 9.70 0.00 982

(Ha, 1998)

10_digit NA NA NA NA (Procter, 1998) 2,3,4,5,6 and

10_digits 74.20 25.80 NA 1,400

2_digit 95.23 4.77 NA 1,000 3_digit 88.01 11.99 NA 1,000 4_digit 80.69 19.31 NA 1,000 5_digit 78.61 21.39 NA 1,000 6_digit 70.49 29.51 NA 1,000

(Lee, 1999)

10_digit NA NA NA NA (Oliveira, 2002) 2_digit 96.88 1,00 2,12 2,370 (MLP-based method) 3_digit 95.38 1,00 3,62 2,385 4_digit 93.38 1,00 5,62 2,345 5_digit 92.40 1,00 6,60 2,316 6_digit 93.12 1,00 5,88 2,169 10_digit 90.24 1,00 8,76 1,217

2_digit 94.81 5.19 0.00 2,370 3_digit 91.61 8.39 0.00 2,385 4_digit 91.25 8.75 0.00 2,345 5_digit 88.30 11.70 0.00 2,316 6_digit 89.07 10.93 0.00 2,169

(Britto, 2003)

10_digit 86.94 13.06 0.00 1,217 (Oliveira, 2004) 2_digit 97.67 2.33 0.00 2,370 (SVM-based method) 3_digit 96.26 3.74 0.00 2,385 4_digit 94.28 5.72 0.00 2,345 5_digit 94.00 6.00 0.00 2,316 6_digit 93.80 6.20 0.00 2,169 10_digit 91.38 8.62 0.00 1,217 (Liu, 2004a) 3_digit 96.82 1,00 2,18 1,476 6_digit 96.74 1,00 2,26 1,471

Document Analysis

12

that enables the use of two sets of features and numeral models: one taking into account both the segmentation and recognition aspects in an implicit segmentation-based process, and another considering just the recognition aspects in a further verification process (see Figure 4). Table 3 summarizes some important contributions on numeral string recognitions.

λλλ 910 ...,,, vcvcvc

Verification

λλλ 910 ...,,, vrvrvr

Re-ranked hypotheses 488 - 0.710, 489 - 0.201, 408 - 0.050,

λλλ 910 ...,,, cccPreprocessing

FFE

SR

FBFE

N Best SR hypotheses 489 - 0.420, 488 - 0.301, 408 - 0.230,

String Contextual-Based

Verification Stage

Fig.4 A two-stage HMM-based method for recognition of numeral strings (Britto,2003) WORD RECOGNITION The majority of research in handwritten word recognition has integrated the lexicon as constraint to build lexicon-driven strategies in opposite to handwritten digit string recognition. The lexicon is a list of possible words that could possibly occur in an image. This lexicon is usually determined by the application at hand. It aims at decreasing the complexity of the problem since the ambiguity makes many characters unidentifible without referring to context. Indeed, the same representation may lead to several interpretations, in the absence of the context which can be a lexicon or grammatical constraints. In Figure 5, for instance, the group of letters in the word “junho” shown inside the circle could be interpreted as “june”, “fune”, “fine”, etc.

Fig.5 Ambiguity in handwritten words

Handwritten word techniques use either analytic or holistic approaches for training and recognition. In an analytic approach, the segmentation of words into segments that relate to characters is required. Nevertheless, this is not a trivial task due to problems such as touching, overlapping, or broken characters. Moreover, this operation is made more diffcult because of the ambiguity encountered in handwritten words. Therefore, most successful


13

analytical methods employ segmentation-based recognition strategies where the segmentation can be explicit (El Yacoubi, 1999)(Arica, 2002)(Kundu, 2002) or implicit (Gillies, 1992)(Cho, 1995)(Mohamed, 1996). In an holistic approach, word recognition is performed considering the whole word. In such a case, there is no attempt to split the word image into segments. Still, it is possible that the image would be segmented in order to produce a sequence of observations. Unlike analytical methods, holistic methods are constrained to applications with a small lexicon size as in bank check processing systems (Dimauro, 1997)(Han, 1997)(Freitas, 2001). Most works in handwritten word recognition assume that the word has been already segmented by an algorithm appropriate to the application domain prior to being processed by word recognizer. Some of them are described as follows. The location and segmentation of handwritten words from their surroundings is a complex task for most real applications. Mohamed and Gader (Mohamed, 1996) combine implicit segmentation-based continuous HMM and explicit segmentation-based dynamic programming (DP) techniques for unconstrained handwritten word recognition. The combination module uses differences in classifer capabilities to achieve significantly better performance. They report results on the BD test set of the CEDAR database that contains 317 city name images. El Yacoubi et al in (El Yacoubi, 1999) design a system to recognize unconstrained handwritten words for large vocabularies. After preprocessing, a word image is divided explicitly into a sequence of segments and then two feature sets are extracted from the sequence of segments. The word models are made up of the concatenation of appropriate letter models and an HMM-based interpolation technique is used to optimally combine the two feature sets. They consider two rejection mechanisms depending on whether or not the word image is guaranteed to belong to the lexicon. The experiments are carried out on 4,313 French city name images manually localized on real mail envelopes. Kim et al in (2000) take an HMM-MLP hybrid system for recognizing cursive script words. They have designed explicit segmentation-based HMM and an holistic approach for the MLP. The main contributions of this work lie in the HMM-based approach and a new multiplication method of combining two distinct classifiers. Experiments are carried out on the handwritten cursive legal words of the CENPARMI database (English set). They use 2,482 word images for testing. Arika et al in (Arika, 2002) propose an analytic scheme, which makes use of a sequence of segmentation and recognition algorithms, for cursive handwriting recognition. First, some global parameters, such as slant angle, baselines, and stroke width and height are estimated. Second, a segmentation method finds character segmentation paths by combining gray scale and binary information. Third, HMM is employed for shape recognition to label and rank the

Document Analysis

14

character candidates. The estimation of feature space information and HMM ranks are combined in a graph optimization problem for word-level recognition. The performance of the system is tested using 2,000 words of the database of Lancester-Oslo/Bergen corpus, which contains single author cursive handwriting. Kapp et al (Kapp, 2004) evaluate the use of the conventional architecture feedforward MLP (multiple layer perceptron) and class-modular for the handwriting recognition (HWR) and also compare the results obtained with previous works in terms of recognition rate. The experiments showed that the class-modular architecture is better than conventional architecture. The obtained average recognition rates were 77.08% using the conventional architecture and 81.75% using the class-modular. This paper also describes a performance study in which a rejection mechanism with multiple thresholds is evaluated for both conventional and class-modular architectures. The multiple thresholds idea is based on the use of N class-related reject thresholds (CRTs). The results indicate that this rejection mechanism can be used appropriately in both architectures. Günter and Bunke (Günter, 2004) combine three classifiers with different architectures for handwritten word recognition. In addition, a new ensemble method working with several base classifiers is applied and the results of the ensemble method are compared to the results of the combination of the three classifiers. An increase of 2.98% was obtained with the best combination scheme. The authors also show that by using classical ensemble methods, such as Bagging and AdaBoost, the performance could also be increased. However, the best performance was achieved with a new ensemble method proposed by the authors, which is distinguished from classical ensemble methods by the fact that it uses several base classifiers, rather than just a single one, to derive an ensemble. The performance of the new ensemble method was 1.5 % higher than the best combination of the base classifiers, 2.94 % higher than the classical ensemble methods, and 4.48 % higher than the best base classifier. Koch et al (Koch, 2004) combine contextual information for recognition of handwritten words extracted from real incoming mail documents. The word recognition process is based on three different sources of information: outputs of a character classifier, contextual information extracted from word shapes and some a priori knowledge. The experiments have shown the benefit of those additional information on the word recognition rates. Koerich et al (Koerich, 2004) propose a fast two–level HMM decoding algorithm to deal with large vocabulary handwriting. The authors propose a non–heuristic fast decoding algorithm which is based on hidden Markov model representation of characters. The decoding algorithm breaks up the computation of word likelihoods into two levels: state level and character level. Given an observation sequence, the two level decoding enables the reuse of character likelihoods to decode all words in the lexicon, avoiding repeated computation of state sequences. In an 80,000–word recognition task, the proposed decoding


15

algorithm is about 15 times faster than a conventional Viterbi algorithm, while maintaining the same recognition accuracy. The performance of some recent applications found in the literature, including those discussed above, are reported in Table 4. The recognition rates are shown for different lexicon sizes at zero-rejection level. In such a case, it is very diffcult to compare the results because of the experiments were conducted on different databases, different classes of words, and different number of testing samples. Generally, the larger the lexicon is, the results get less satisfactory. Besides the lexicon size, the results also depend on the ambiguity among the classes of words being considered and whether a system is writer-dependent or not (omni-writer). However, the results are helpful in order to illustrate the current state of art on word recognition.

Table 4 Performance of some recent applications found in the literature

Method Classifier Lexicon Size

RR (%)

Test Set Database Coments

(Bunke, 1995) HMM 150 98.4 3,000 Words (ENG) CUR, WD (Mohamed, 1996) HMM-DP 100 89.3 317 City names (Eng) UNC OMNI (Knerr, 1998) HMM-NN 30 92.9 40,000 LA words (FR) CUR, OMNI (Guillevic,1998) HMM-k-NN 30 86.7 4,500 LA words (ENG) CUR, OMNI (El Yacoulbi, 1999) HMM 100 96.3 4,313 City names (FR) UNC OMNI (El Yacoulbi, 1999) HMM 1,000 88.9 4,313 City names (FR) UNC, OMNI (Kim, 2000) HMM-MLP 32 92.2 2,482 LA words (ENG) CUR, OMNI (Freitas, 2001) HMM 39 77.0 2,387 LA words (POR) UNC, OMNI (Oliveira Jr, 2002) MLP 12 87.2 1,200 Month words

(POR) UNC, OMNI

(Xu, 2002) HMM-MLP 29 85.3 2,063 Month words (FR/ENG)

UNC, OMNI

(Kundu , 2002) HMM 100 88.2 3,000 Postal words UNC OMNI (Arika, 2002) HMM 1,000 90.8 2,000 Words (ENG) CUR, WD (Kapp, 2004) Class-

modular MLP 39 81.7 2,387 LA words (POR) UNC, OMNI

(Kock, 2004) MLP 1,000 67,8 500 Letters (FR) UNC OMNI (Koerich, 2004) HMM 1,000 91,0 4,674 City names (FR) UNC OMNI

LA:Legal Amount, FR:French, POR:Portuguese, ENG:English UNC:Unconstrained,CUR:Cursive, OMNI:Omni-writer, WD:Writer-Dependent SENTENCE RECOGNITION

Recognition of isolated characters, numeral strings and words has been extensively studied in the literature. However, only recently the recognition of a sentence composed of a

Document Analysis

16

sequence of words or different data types has been investigated. Some typical applications on sentence recognition are reading texts from pages (Favata, 1998)(Marti, 2001a)(Marti, 2001b), street names from postal addresses (G. Kim, 1998)(Park, 1999)(Park, 2002)(El Yacoubi, 2002), processing of legal amounts (Kaufmann, 2000)(Gorski, 2001) and dates (Suen, 1999b)(Xu, 2002)(Morita, 2003) on checks. In such applications, sentences are segmented into its constituent parts and make use of word and numeral string recognition techniques. In the literature two main different strategies of segmentation can be observed: segmentation-based and segmentation-free methods. In the segmentation-based strategy, the most frequently used method splits explicitly a sentence into its parts usually based on spatial distance clues, i.e., the distances (gaps) between adjacent components (Mahadevan, 1995)(Favata, 1998)(Marti, 2001a). Notwithstanding, the segmentation based on gap metrics shows its limits rapidly when handwritten sentences do not have a uniform spacing. Besides, the gaps between components can not be easily estimated by a 1-dimensional metric. In order to overcome the foregoing problem, Kim et al in (G. Kim, 1998) propose an intelligent word segmentation method applied to street name images. This method incorporates the author's writing style in terms of spacing using an NN. The NN has been designed to split a sentence into segments and then an exhaustive combination of word segments is submitted to a word recognizer with given a lexicon. Statistical characteristics of character segmentation are used to limit the number of combinations. Further improvements of this method were given in (Park, 1999). In the same vein, Xu et al in (Xu, 2001) make use of different kinds of knowledge at different segmentation stages. The knowledge includes information on the writing style by using an NN to differentiate between numerical and alphabetical data, and syntactical and semantic constraints. In this strategy, multiple hypotheses of segmentation are often considered in order to improve the performance of segmentation. The idea is to generate a list of segmentation hypotheses where each word candidate of each sentence hypothesis is submitted to recognition. The result of segmentation and linguistic analysis can be used to rank each sentence hypothesis. Favata et al (Favata, 1998) propose a two step approach to sentence recognition. After a sentence has been divided into segments, they group segments as a function of their spatial inter-relationship and then as a function of their response to recognizer. The output of the recognizer is a directed graph which contains multiple interpretations of the sentence. The second step searches this graph with a linguistic processor which ranks each path according to statistics of the language and certain measured statistical characteristics of the recognizer. On the other hand, the segmentation-free strategy treats complete handwritten sentences as single units. Scagliola (Scagliola, 1998) proposes a system for recognizing legal amounts on Italian bank checks without word segmentation. The author demonstrates that when the a


17

priori knowledge is described by a regular grammar, the phrase recognition can be performed by a simple extension of the dynamic programming algorithms used for single word recognition. Marti et al (Marti, 2000) present a system for reading unconstrained handwritten text (see Figure 6). The kernel of the system is an HMM for handwriting recognition. This HMM is enhanced by a statistical language model incorporating linguistic information beyond the word level. The HMM has a hierarchical structure with character models at the lowest level. These models are concatenated to words and to whole sentences. Under such an architecture, the segmentation of a text line into individual words is not required, instead, it is obtained during the recognition process. In the same manner, El Yacoubi et al (El Yacoubi, 1999) build an embedded HMM network connecting street name HMMs with extraneous HMMs in order to conjointly locate and recognize street name without the need of an a priori segmentation of the street line.

Fig.6 Scheme proposed in (Marti, 2000):

a) Recognition network based on HMMs; b) image of a corrected line fragment Zimmermann et al (Zimmermann, 2004) present another important contribution. The authors investigate the impact of bigram and trigram language models on the performance of a Hidden Markov Model (HMM) based offline recognition system for handwritten sentences. The language models are trained on the LOB corpus which is supplemented by various additional sources of text, including sentences from additional corpora and random sentences produced by a stochastic context-free grammar (SCFG). Vinciarelli et al (Vinciarelli, 2004) propose a system for offline recognition of large vocabulary unconstrained handwritten text. They use statistical language models in order to

Document Analysis

18

improve the system performance. The use of language models has produced an improvement to the accuracy of the system. The experimental results considering a lexicon of 50,000 words have shown a reduction of 50% on error rate for single writer data and 25% for multiple writer data. Table 5 reports the performance of some methods on sentence recognition found in the literature. In such cases, part of the recognitions errors can come from the segmentation stage.

Table 5 Performance of some methods on sentence recognition

Sentence level Word level Method Application RR

(%) Error (%)

Test set

RR (%)

Lexicon size

Test set

(Kaufmann, 2000) Legal amount (GER) 71.9 8.1 1,500 - - - (Marti, 2000) Text lines (Eng) - - - 79.5 412 4,523 (Marti, 2001a) Text lines (Eng) - - - 73.3 412 3,899 (Marti, 2001b) Text lines (Eng) - - - 61.8 2,703 11,000 (El Yaboulbi, 2002) Street names (FR) 45.6 1.5 2,836 - - - (Zimmermann, 2004) Text lines (Eng) 14.0 - 200 81.8 8,819 - (Vinciarelli, 2004) Text lines (Eng) - - - 40.0 10,000 -

TOWARDS PERFORMANCE Feature Selection An important issue in constructing classifiers is the selection of the best discriminative features. In many applications of handwriting recognition, it is not unusual to find problems involving hundreds of features. However, it has been observed that beyond a certain point, the inclusion of additional features leads to a worse rather than better performance. Moreover, the choice of features to represent the patterns affects several aspects of the pattern recognition problem such as accuracy, required learning time, and the necessary number of samples. This apparent paradox presents us with a feature selection problem in automatic design of pattern classifiers. Such a problem refers to the task of identifying and selecting an effective subset of features to represent patterns from a larger set of often mutually redundant or even irrelevant features. This is not a trivial problem since features are seldom entirely independent. There may be redundancy, where certain features are correlated so that it is not necessary to include all of them in modeling, and interdependence, where two or more features between them convey important information that is obscure if any of them is included on its own. Recently researchers have applied feature selection techniques to reduce the complexity of classifiers and improve performance as well (G. Kim, 2000)(Oliveira, 2003a)(Nunes, 2004).


19

Multiple Classifiers In the recent years, combination of classifiers has been attracted theoretical and practical attention. In the long run, the combined decision is supposed to be better (more reliable) than the classification decision of the best individual classifier. Such an idea appears under a variety of names in the literature: classifier fusion (Gader, 1996b), classifier combination (Kittler, 1998)(Xu, 1992), mixture of experts (Jacobs, 1991), committees (Bishop, 1995), etc. The literature on theoretical classifier combination has grown rapidly (Xu, 1992)(Alkoot, 1999)(Kittler, 1998)(Tax, 2000)(Kuncheva, 2002). The papers advocate different fusion strategies, demonstrate the benefits of classifier, and provide a theoretical underpinning of the various strategies commonly used in multiple expert fusion. Some interesting results of combination scheme are also reported in practical applications of handwriting recognition. Suen et al (Suen, 1992) combined four structural methods using variants of majority vote. High performance was reported on CENPARMI database. Xu et al (Xu, 1992) used the same four classifiers and results are provided using Dempster-Shafter and Bayesian formalism. Lam and Suen (Lam, 1995) applied weighted voting to handwritten digit recognition using seven classifiers. Kittler et al (Kittler, 1998) applied several classifier combination strategies such as product rule, sum rule, min rule, max rule, median rule, and majority vote to recognize handwritten numerals. Gorski (Gorski, 1996) demonstrated that for large databases, in the context of bank cheque recognition, neural networks provide compelling results as a classifier integrator. More recently, ensemble of classifiers seems to be the new trend in the fields of machine learning and pattern recognition. Ensemble of classifiers are characterized by the fact they produce several classifiers out one given base classifiers. It can be generated by changing the training set (Breiman, 1996), the input features (Ho, 1998)(Optiz, 1999), the input data by injecting randomness (Valentini, 2002) or still the parameters and the architecture of the base classifiers (Partridge, 1996). In the field of handwriting recognition some authors have been used ensemble of classifiers in order to improve both accuracy and reliability of the systems. Nishimura et al (Nishimura, 1999) have used bagging to improve the performance of an HMM-based handwritten character recognition system. Gunter and Bunke (Gunter, 2003) apply several different method such as bagging, boosting, random subspace, and architecture variation to generate ensembles of classifiers in the context of handwritten word recognition. They show through experimentation that even for a intensively optimized base classifier, it is possible to get some improvement. Oliveira et al (Oliveira, 2003b) propose a strategy for ensemble generation based on feature selection, which takes into account a multi-objective genetic algorithm. They have demonstrated that the reliability of the classifiers (neural networks in this case) can be improved considerably by using ensemble of classifiers. Likewise, Morita et al (Morita, 2004) have applied a similar strategy for creation of ensembles of Markovian

Document Analysis

20

classifiers in the context of unsupervised feature selection. Compelling improvements were reported in the problem of handwritten month word recognition. Liu and Narukawa (Liu, 2004b) have devised 14 normalization methods to build the ensemble. They have shown through experimentation that such normalization methods are complementary and thus the ensemble is able to improve the recognition accuracy. Other strategy that can increase the recognition rate in a relatively easy way with a small additional cost is through the use of verification. Such a scheme consists of refining the top few candidates in order to enhance the recognition rate economically. Such a kind of scheme has been successfully applied to handwriting recognition in (Takahashi, 1993)(Zhou, 2000)(Britto, 2002)(Oliveira, 2003c). A different multiple classifier system has been emerged in the recent years in the field of handwriting recognition. It consists in developing two kind of classifiers: cheap and complex. The cheap ones are used in a first level to classify easy cases, while the complex ones are used to classify what is rejected by the previous level. Vuurpijl et al (L. Vuurpijl, 2003) applied this strategy to classify handwritten characters. In that case, they have used the expensive SVM in the second level. A similar approach was adopted by Milgram et al (Milgram, 2004). They have focused in the problem of handwritten digits to demonstrated that it is possible to reduce the complexity of the recognition task using this kind of strategy. In the save vein, Nunes et al (Nunes, 2004) have proposed six levels of neural networks to classify handwritten digits. The classifiers were generated based on feature selection, i.e., the first classifier uses few features, therefore less complex, while the last one uses the entire feature set. Synthetic Data The availability of large amount of data is fundamental for building a reliable handwriting recognition system. The acquisition of large database has become an issue in the community of handwriting recognition. A survey convering the main databases until 1996 can be found in (Guyon, 1997). After that, database with complete handwritten sentences were proposed in (Senior, 1998)(Marti, 2002). The database reported in (Marti, 2002) is significantly larger than (Senior, 1998) and include text from multiple writers. Since collecting data and preparing its ground truth are very expensive and time consuming tasks, some researchers propose to use synthetic data for training. In (Ha, 1997)(Mori, 2000) the authors describe the synthetic generation of isolated characters. The synthetic generation of handwritten words and sentences has been described in (Guyon, 1996) and (Helmers, 2003) . In an upper level, Varga and Bunke (Varga, 2003) proposed an geometrical distortion model for complete lines of handwritten text. They have demonstrated through several experiments using HMM-based classifiers that the use of synthetic data can improve the


21

recognition performance, particularly when the training data has been produced by a small population of writers. CONCLUSION

We agree when the author, in (Lorette, 1999), say that the ultimate goal of handwriting recognition is to have machines which can read any text with the same recognition accuracy as humans but at a faster rate. In fact, we hope, in the near future, writing will serve not only for communication among humans but also it will serve for communication of humans and machines. The DAR research has shown significant contributions in this direction. However, more research is needed to deal with the following bottlenecks:

• Character recognition: it is clear that the maturity was reached just for isolated digits. More research is needed, however, to improve the recognition rates for alphabetic characters. Few researches have addressed this complicated subject, in which there is a larger number of ambiguous alphabetic characters, multiple patterns to represent a single character (upper and lowercase), cursive representation of letters, and disconnected and multi–stroke characters. Suggestions: feature extraction and selection, ensemble of classifiers.

• Numeral string recognition: the main bottleneck is the segmentation step. Segmentation-based methods reach better accuracy, however they are usually time consuming. On the other hand, segmentation-free usually shows some loss in terms of recognition rates when segmentation and recognition are contemplated in the same process. Suggestions: creation of a database with touching digits, segmentation based on an intelligent process in order to reduce the segmentation-path candidates, post-processing techniques.

• Word recognition: it seems that the problem of word recognition in the case of small and static lexicons is essentially solved. However, the accuracy drops from 90% for volabularies of hundreds of words to less than 80% for vocabularies of ten thousands of words (Koerich, 2004). The negative effects are also observed on the recognition speed. While the recognition time is on the order of tens of seconds for small vocabularies, it increases to few minutes for very large vocabularies. Suggestions: the study of fast decoding algorithms, feature extraction and selection, ensemble of classifiers, post-processing techniques.

• Sentence recognition: we can find in the literature some success approaches to deal with this problem, but only for small and specific domains, such as: address or check reading. The challenge is to recognize unconstrained text. Again, the main bottleneck is the segmentation step usually used to segment complete sentences into single words. Suggestions: recognition of the entire sentence based on language models (bigrams and trigrams), post-processing techniques.

Document Analysis

22

Finally, one can see that most of the time the approaches proposed for hadwriting recognition are evaluated on noiseless documents. Thus, the DAR researches have a big challenge to reach the ultimate objective of making computers approach human performance in terms of understanding documents. REFERENCES

[1] K. Jain, R. P. W. Duin, J. Mao (2000). Statistical pattern recognition: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1):4-37.

[2] P. D. Gader, B. Forester, M. Ganzberger, A. Billies, B. Mitchell, M. Whalen, T. Youcum (1991). Recognition of handwritten digits using template and model matching. Pattern Recognition, 5(24):421-431.

[3] G. Dimauro, S. Impedovo, G. Pirlo, A. Salzo (1997). Automatic bankcheck processing: A new engineered system. In S.Impedovo et al, editor, International Journal of Pattern Recognition and Artificial Intelligence, World Scientific, pp. 467-503.

[4] S. L. Xie, M. Suk (1988). On machine recognition of hand-printed chinese character by feature relaxation. Pattern Recognition, 21(1):1-7.

[5] D. Guillevic, C. Y. Suen (1995). Cursive script recognition applied to the processing of bank cheques. In Proc. of 3th International Conference on Document Analysis and Recognition, Montreal-Canada, August, pp. 11-14.

[6] L. Mico, J. Oncina (1999). Comparison of fast nearest neighbour classifier for handwritten character recogniton. Pattern Recognition Letters, 19(3-4):351-356.

[7] R. O. Duda, P. E. Hart, D. G. Stork (2001). Pattern Classification. John Wiley and Sons, second edition edition.

[8] J. Schurmann (1996). Pattern Classification - A unified view of statistical and neural approaches. Wiley interscience.

[9] L. R. Rabiner (1989). A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition. Proceedings of the IEEE, Vol. 77, No. 2, pp.257-286.

[10] El Yacoubi, M. Gilloux, R. Sabourin, C. Y. Suen (1999). An hmm-based approach for off-line unconstrained handwritten word modeling and recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(8):752-760.

[11] Kundu, Y. He, M. Chen (2002). Alternatives to variable duration hmm in handwriting recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(11):1275-1280.

[12] W. Senior, A. J. Robinson (2002). An off-line cursive handwriting recognition system. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):309-321.

[13] F. Grandidier (2003). Un Nouvel Algorithme de Sélection de Caractéristiques- Application à la Lecture Automatique de l'ecriture Manuscrite. PhD thesis, École de Technologie Supérieure, Montreal-Canada, Janvier.

[14] J. Cai and Z. Q. Liu (1999). Integration of structural and statistical information for unconstrained handwritten numeral recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(3):263-270.


23

[15] Britto Jr., R. Sabourin, F. Bortolozzi, and C. Y. Suen (2001). A two-stage hmm based system for recognizing handwritten numeral strings. In Proc. 6th International Conference on Document Analysis and Recognition, Seattle-USA, September, pp. 396-400.

[16] S. I. Abuhaiba, P. Ahmed (1993). A fuzzy graph theoretic approach to recognize the totally unconstrained handwritten numerals. Pattern Recognition, 26(9):1335-1350.

[17] P. D. Gader, J. M. Keller, J. Cai (1995). A fuzzy logic system for detection and recognition of street number fields on handwritten postal addresses. IEEE Transactions on Fuzzy Systems, 3(1):83-95.

[18] Lazzerini, F. Marcelloni (2000). A linguistic fuzzy recognizer of off-line handwrriten characters. Pattern Recognition Letters, 21(4):319-327.

[19] V. Vapnik (1995). The Nature of Statistical Learning Theory. Springer-Verlag, New York-USA.

[20] N. E. Ayat, M. Cheriet, C. Y. Suen (2002). Optimization of the SVM kernels using an empirical error minimization scheme. In Proc. of the International Workshop on Pattern Recognition with Support Vector Machine, Niagara Falls-Canada, August, pp. 354-369.

[21] H. Byun, S. W. Lee (2002). Applications of support vector machines for pattern recognition. In Proc. of the International Workshop on Pattern Recognition with Support Vector Machine, Niagara Falls-Canada, August, pp. 213-236.

[22] L. S. Oliveira, R. Sabourin (2004). Support Vector Machines for Handwritten Numerical String Recognition, 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 39-44.

[23] M. Shridhar and A. Badreldin (1986). Recognition of isolated and simply connected handwritten numerals. Pattern Recognition, 19(1):1-12.

[24] H. Y. Kim and J. h. Kim (1998). Handwritten korean character recognition based on hierarchical randon graph modeling. In Proc. 6th International Workshop on Frontiers of Handwriting Recognition, Taegon-Korea, August, pp. 557-586.

[25] M. Bishop (1995). Neural Networks for Pattern Recognition. Oxford Univ. Press, Oxford-U.K.

[26] Y. LeCun, L. Bottou, G. B. Orr, K. R. Muller (1998a). Eficient backprop. In G. Orr and K. Miller, editors, Neural Networks: Tricks of the Trade. Springer.

[27] G. P. Zhang (2000). Neural networks for classification: a survey. IEEE Transactions on Systems, Man, and Cybernetics - Part C: Applications and Reviews, 30(4):451-462.

[28] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner (1998b). Gradient-based learning applied to document recognition. Proc. of IEEE, 86(11):2278-2324, 1998.

[29] Zhang, M. Fu, H. Yan, and M. A. Fabri (1999). Handwritten digit recognition by adaptative-subspace self organizing map. IEEE Trans. on Neural Networks, 10:939-945.

[30] O. Matan, J. C. Burges, Y. LeCun, J. S. Denker (1992). Multi-digit recognition using a space displacement neural network. In J. E. Moody, S. J. Hanson, and 165 R. L. Lippmann, editors, Advances in Neural Information Processing Systems, volume 4, Morgan Kaufmann, pp. 488-495.

[31] Lethelier, M. Leroux, M. Gilloux (1995). An automatic reading system for handwritten numeral amounts on french checks. In Proc. 3th International Conference on Document Analysis and Recognition, Montreal-Canada, August, pp. pages 92-97.

[32] J. Zhou (1999). Recognition and Verification of Unconstrained Handwritten Numeral. PhD thesis, Concordia University, Montreal-Canada, November.

Document Analysis

24

[33] Ling, M. Lizaraga, N. Gomes, A. Koerich (1997). A prototype for brazilian bankcheck recognition. In S.Impedovo et al, editor, International Journal of Pattern Recognition and Artificial Intelligence,World Scientific, pp. 549-569.

[34] J. Kittler, M. Hatef, R. Duin, J. Matas (1998). On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-239.

[35] Q. Xu (2002). Automatic Segmentation and Recognition System for Handwritten Dates on Cheques. PhD thesis, Concordia University, Montreal-Canada, December.

[36] L. Koerich (2002). Large Vocabulary Off-Line Handwritten Word Recognition. PhD thesis, École de Technologie Supérieure, Montreal-Canada, August.

[37] S. Oliveira, R. Sabourin, F. Bortolozzi, and C. Y. Suen (2002). Automatic recognition of handwritten numerical strings: A recognition and verification strategy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(11):1438-1454.

[38] O. D. Trier, A.K. Jain, T. Taxt (1996). Feature Extraction Methods for Character recognition - a Survey. Pattern Recognition, Vol 29, No. 4, pp. 641-662.

[39] S. Loncaric (1998). A Survey of Shape Analysis Techniques. Pattern Recognition, Vol. 31, No. 8, pp. 983-1001.

[40] Kimura, M. Shridhar (1992). Segmentation-recognition algorithm for handwritten numeral strings. Machine Vision Applications, No. 5, pp. 199-210.

[41] P. D. Gader and M. A. Khabou (1996a). Automatic Feature Generation for Handwritten Digit Recognition. IEEE Transactions on Pattern Analysis and Machine Intellingence, Vol. 18, No. 12, pp. 1256-1261.

[42] K. Cheung, D. Yeung (1998). A Bayesian Framework for Deformable Pattern Recognition with Application to Handwritten Character Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20, No. 12, December, pp.1382-1388.

[43] T. Hirano, Y. Okada, F. Yoda (1997). Structural Character Recognition Using Simulated Annealing. Proceedings of the Fourth International Conference on Document Analysis and Recognition, Vol. 2, pp. 507-510.

[44] J. Cai, Z. Liu (1998). Integration of Structural and Statistical Information for Unconstrained Handwritten Numeral Recognition. Proceedings of the Fourteenth International Conference on Pattern Recognition, Vol. I, pp. 378-380.

[45] Heute L., T. Paquet, J. V. Moreau, Y. Lecourtier, C. Olivier (1998). A structural/statistical feature-based vector for handwritten character recognition. Pattern Recognition Letters, 19, pp. 629-641.

[46] S. Britto Jr., R. Sabourin, F. Bortolozzi, C. Y. Suen (2004). Foreground and Background Information in an HMM-Based Method for Recognition of Isolated Characters and Numeral Strings, 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, 2004, Kokubunji, Tokyo, Japan, pp 371-376.

[47] H. Park, S. Lee (1998). A Truly 2-D Hidden Markov Model for Off-Line Handwritten Character Recognition. Pattern Recognition, Vol. 31, No. 12, pp. 1849-1864.

[48] Zhang, M. Fu, H. Yan (1998). A Modular Classification Scheme with Elastic Net Models for Handwritten Digit Recognition. Proceedings of the Fourteenth International Conference on Pattern Recognition, pp. 1859-1861.

[49] C.Y Suen, K. Liu, N.W. Strathy (1999a). Sorting and recognizing cheques and financial documents. In: S. W. Lee, Y. Nakano (Eds.), Document Analysis Systems: Theory and Practice, Springer, Belin, pp. 173-187.


25

[50] C-L. Liu, H. Sako, H. Fujisawa (2004a). Effects of classifier structures and training regimes on integrated segmentation and recognition of handwritten numeral strings, IEEE Trans on Patterns Analysis and Machine Intelligence, 26(11):1395-1407.

[51] L.-N. Teow, K.-F Loe (2002). Robust vision-based features and classification schemes for off-line handwritten digit recognition. Pattern Recognition 35 (11), pp. 2355-2364.

[52] J. Milgram, M. Cheriet, R. Sabourin (2004). Speeding up the decision making of Support Vector Classifier. 9th Intern. Workshop on Frontiers in Handwriting Recognition, pp. 57-62.

[53] J. Zhou, G. Qiang, C. Y. Suen (1997). A High Performance Hand-printed Numeral Recognition System with verification Module. Proceedings of the Fourth International Conference on Document Analysis and Recognition, Vol. 2, pp. 293-297.

[54] S. Belongie, J. Malik, J. Puzicha (2002). Shape matching and object recognition using shape contexts. IEEE Trans. Pattern Analysis and Machine Intelligence, 24 (2), pp. 509-522.

[55] C-L. Liu, K. Nakashima, H. Sako, H. Fujisawa (2003). Handwritten digit recognition: benchmarking of state-of-the-art techniques. Pattern Recognition 36, pp. 2271-2285.

[56] T. Ha, H. Bunke (1997). Off-line handwritten numeral recognition by perturbation method. IEEE Trans. Pattern Analysis and Machine Intelligence, 19 (5), pp. 535-539.

[57] V. Delevski, S. Stankovic (1998). Recognition on Handwritten Digits based on their Topological and Morphological Properties. Advances in Pattern Recognition, Joint IAPR International Workshop SSPR’98 and SPR’98, pp. 516-523.

[58] C.-L. Liu, M. Nakagawa (1999). Handwritten numeral recognition using neural networks: improving the accuracy by discriminative training. Proc. of the Fifth International Conference on Document Analysis and Recognition, pp. 257-260.

[59] J. X. Dong, A. Krzyzak, C.Y. Suen (2001a). A muti-net learning framework for pattern recognition. Proc. of the Sixth International Conference on Document Analysis and Recognition, Seattle, pp. 328-332.

[60] C.-L. Liu, K. Nakashima, H. Sako, and H. Fujisawa (2002). Handwritten digit recognition using state-of-the-art techniques. In Proc. of 8th International Workshop on Frontiers of Handwriting Recognition. pp. 320–325.

[61] DeCoste and B. Schölkopf (2002). Training invariant support vector machines. Machine Learning Journal, 46(1-3):161-190.

[62] S. Oh, C. Y. Suen (1998). “Distance features for neural network–based recognition of handwritten characters”. International Journal on Document Analysis and Recognition, 1(2):73-88.

[63] J. X. Dong, A. Krzyzak, C. Y. Suen (2001b). Local learning framework for recognition of lowercase handwritten characters. In Proc. Int. Workshop on Machine Learning and Data Mining in Pattern Recognition, Leipizig, Germany, pp. 226-238.

[64] Yu, H. Yan (2001). Separation of touching handwritten multi-numeral strings based on morphological structural features. Pattern Recognition, 34:587-599.

[65] K. K. Kim, J. H. Kim, C. Y. Suen (2002). Segmentation-based recognition of handwritten touching pairs of digits using structural features. Pattern Recognition, 23(1):13-21.

[66] Z. Lu, Z. Chi, W. Siu, P. Shi (1999). A background-thinning based approach for separating and recognizing connected handwritten digit strings. Pattern Recognition, 32:921-933.

[67] Y. K. Chen, J. F. Wang (2000). Segmentation of single- or multiple-touching handwritten numeral string using background and foregound analysis. IEEE Trans. on PAMI, 22(11):1304-1317.

Document Analysis

26

[68] U. Pal, A. Belaid, Ch. Choisy (2003). Touching numeral segmentation using water reservoir concept. Pattern Recognition Letters, 24:261-272.

[69] B. Bose, S-S Kuo (1994). Connected and Degraded Text Recognition using Hidden Markov Model. Pattern Recognition, Vol. 27, No. 10, pp. 1345-1363.

[70] J. Elms, S. Procter, J. Illingworth (1998). The Advantage of using an HMM-based Approach for Faxed Word Recognition. International Journal on Document Analysis and Recognition, pp. 18-36.

[71] S. Procter, A. J. Elms (1998). The recognition of handwritten digit strings of unknown length using hidden Markov models. Proceedings of the Fourteenth International Conference on Pattern Recognition, pp. 1515-1517.

[72] S. Britto Jr. R. Sabourin, F. Bortolozzi, C. Y. Suen (2003). The Recognition of Handwritten Numeral Strings Using a Two-stage HMM-based Method. International Journal on Document Analysis and Recognition, 5:102-117.

[73] L. Matin, M. Rashid, J. A. Pittman (1993). Integrated segmentation and recognition through exhaustive scans or learned saccadic jumps. International Journal of Pattern Recognition and Artificial Intelligence. Vol. 7, No. 4, pp. 831-847.

[74] T. M. Ha, M. Zimmermann, H. Bunke (1998). Off-Line Handwritten Numeral String Recognition by Combining Segmentation-Based and Segmentation-Free Methods. Pattern Recognition, Vol. 31, No. 3, pp. 257-272.

[75] S.-W Lee, S-Y Kim (1999). Integrated Segmentation and Recognition of Handwritten Numerals with Cascade Neural Network. IEEE Transactions on Systems, Man, and Cybernetics – Part C: Applications and Reviews, Vol. 29, No. 2, pp. 285-290, February.

[76] Arika, F. T. Yarman-Vural (2002). Optical character recognition for cursive handwriting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(6):801-813, 2002.

[77] M. Gillies (1992). Cursive word recognition using hidden markov models. In Proc. Fifth U.S. Postal Service Advanced Technology Conference, pages 557-562.

[78] W. Cho, S. W. Lee, J. H. Kim (1995). Modeling and recognition of cursive words with hidden markov models. Pattern Recognition, 28(12):1941-1953.

[79] A. Mohamed and P. Gader (1996). Handwritten word recognition using segmentation-free hidden markov modeling and segmentation-based dynamic programming techniques. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(5):548-554.

[80] K. Han and I. K. Sethi (1997). An off-line cursive handwritten word recognition system and its aplication to legal amount interpretation. In S.Impedovo et al, editor, International Journal of Pattern Recognition and Artiffcial Intelligence, World Scientific, pp. 757-770.

[81] Freitas, F. Bortolozzi, and R. Sabourin (2001). Handwritten isolated word recognition: An approach based on mutual information for feature set validation. In Proc. 6th International Conference on Document Analysis and Recognition, Seattle-USA, September, pp. 665-669.

[82] J. H. Kim, K. K. Kim, C. Y. Suen (2000). An HMM-MLP hybrid model for cursive script recognition. Pattern Analysis and Applications, 3:314-324.

[83] M. N. Kapp, C. Freitas, R. Sabourin (2004). Handwritten Brazilian Month Recognition: An Analysis of Two NN Architectures and a Rejection Mechanism, 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 209-214.


27

[84] S. Günter, H. Bunke (2004). Combination of three classifiers with different architectures for handwritten word recognition. 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 63-68.

[85] Koch, T. Paquet, L. Heutte (2004). Combination of Contextual Information for Handwritten Word Recognition. 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 468-473.

[86] L. Koerich, R. Sabourin, C. Y. Suen (2004) Fast Two–Level HMM Decoding Algorithm for Large Vocabulary Handwriting Recognition, 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 232-238.

[87] Bunke, M. Roth, and E. G. Schukat-Talamazzini (1995). Off-line cursive handwriting recognition using hidden markov models. Pattern Recognition, 28(9):1399-1413.

[88] S. Knerr and E. Augustin (1998). A neural network-hidden markov model hybrid for cursive word recognition. In Proc. of 14th International Conference on Pattern Recognition, volume 2, Brisbane-Australia, August, pp. 1518-1520.

[89] Guillevic and C. Y. Suen (1998). HMM-KNN word recognition engine for bank cheque processing. In Proc. of 14th International Conference on Pattern Recognition, Brisbane-Australia, August, pp 1526-1529.

[90] H. Kim, K. K. Kim, and C. Y. Suen (2000). An HMM-MLP hybrid model for cursive script recognition. Pattern Analysis and Applications, 3:314-324.

[91] J. de Oliveira Jr., J. M. de Carvalho, C. Freitas, and R. Sabourin (2002). Feature sets evaluation for handwritten word recognition. In Proc. 8th International Workshop on Frontiers of Handwriting Recognition, Niagara on the Lake-CA, August, pp. 446-451.

[92] Q. Xu, J. H. Kim, L. Lam, and C. Y. Suen (2002). Recognition of handwritten month words on bank cheques. In Proc. 8th International Workshop on Frontiers of Handwriting Recognition, Niagara on the Lake-CA, 169 August, pp. 111-116.

[93] J. T. Favata, S. N. Srihari, V. Govindaraju (1998). Off-line handwritten sentence recognition. In Proc. 6th International Workshop on Frontiers of Handwriting Recognition, Taegon-Korea, August, pp. 171-176.

[94] U. Marti, H. Bunke (2001a). Text line segmentation and word recognition in a system for general writer independent handwriting recognition. In Proc. 6th International Conference on Document Analysis and Recognition, Seattle-USA, September, pp. 159-163,

[95] U. Marti, H. Bunke (2001b). Using a statistical language model to improve the performance of an hmm-based cursive handwriting recognition system. International Journal of Pattern Recognition and Artificial Intelligence, 15(1):65-90.

[96] Kim, V. Govindaraju (1998). Handwritten phrase recognition as applied to street name images. Pattern Recognition, 31(1):41-51.

[97] J. Park, V. Govindaraju, S. N. Srihari (1999). Eficient word segmentation driven by unconstrained handwritten phrase recognition. In Proc. 5th International Conference on Document Analysis and Recognition, Bangalore-India, September, pp. 605-608.

[98] J. Park, V. Govindaraju (2002). Use of adaptive segmentation in handwritten phrase recognition. Pattern Recognition, 35:245-252.

[99] El Yacoubi, M. Gilloux, and J. M. Bertille (2002). A statistical approach for phrase location and recognition within a text line: An application to street name recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(2):172-188.

Document Analysis

28

[100] Kaufmann, H. Bunke (2000). Automated reading of cheque amounts. Pattern Analysis and Applications, 3:132-141.

[101] Gorski, V. Anisimov, E. Augustin, O. Baret, S. Maximov (2001). Industrial bank chech processing: the A2iA checkreader. International Journal on Document Analysis and Recognition, 3:196-206, 2001.

[102] Y. Suen, Q. Xu, and L. Lam (1999b). Automatic recognition of handwritten data on cheques - fact or fiction? Pattern Recognition Letters, 20(13):1287-1295, 1999.

[103] Q. Xu, L. Lam, and C. Y. Suen (2001). A knowledge-based segmentation system for handwritten dates on bank cheques. In Proc. 6th International Conference on Document Analysis and Recognition, Seattle-USA, September, pp. 384-388.

[104] Morita, R. Sabourin, F. Bortolozzi, C. Y. Suen (2003). Segmentation and Recognition of Handwritten Dates: An HMM-MLP Hybrid Approach. International Journal on Document Analysis and Recognition, 6(4):248-262,

[105] U. Mahadevan and R. C. Nagabushnam (1995). Gaps matrices for word separation in handwritten lines. In Proc. 3th International Conference on Document Analysis and Recognition, Montreal-Canada, August, pp. pages 124-127.

[106] Scagliola (1998). Search algorithms for the recognition of cursive phrases without word segmentation. In Proc. 6th International Workshop on Frontiers of Handwriting Recognition, Taegon-Korea, August 1998, pp. 123-132.

[107] U.-V. Marti, H. Bunke (2000). Handwritten Sentence Recognition. Proc. of the International Conference on Pattern Recognition, Barcelona, Spain, vol. 3, pp. 3467-3470.

[108] M. Zimmermann, H. Bunke (2004) N-Gram Language Models for Offline Handwritten Text Recognition. 9th International Workshop on Frontiers in Handwriting Recognition, October 26-29, Kokubunji, Tokyo, Japan, pp 203-208.

[109] Vinciarelli, S. Bengio, H. Bunke (2004). Offline Recognition of Unconstrained Handwritten Texts Using HMMs and Statistical Language Models. IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 26, number 6, pp 709-720.

[110] G. Kim, S. Kim (2000). Feature selection using genetic algorithms for handwritten character recognition. 7th Int. Workshop on Frontiers in Handwriting Recognition, pages 103-112.

[111] L. S. Oliveira, R. Sabourin, F. Bortolozzi, C. Y. Suen. (2003a). A methodology for feature selection using multi-objective genetic algorithms for handwritten digit string recognition. Int. Journal on Pattern Recognition and Artificial Intelligence, 17(6):903-930.

[112] C. Numes, A. Britto Jr, C. Kaestner, R. Sabourin (2004). An optimized hill climbing algorithm for feature subset selection: Evaluation on Handwritten Character Recognition, 9th International Workshop on Frontiers in Handwriting Recognition, pp. 365-370.

[113] D. Gader, M. A. Mohamed, J. M. Keller (1996b). Fusion of handwritten word classifiers. Pattern Recognition Letters, 17(6):577-584.

[114] Kittler, M. Hatef, R. Duin, J. Matas (1998). On combining classifiers. IEEE Trans. on Pattern Analysis and Machine Intelligence, 20(3):226-239.

[115] L. Xu, A. Krzyzak, C. Y. Suen (1992). Methods of combining multiple classifiers and their applications to handwriting recognition. IEEE Trans. on SMC, 22(3):418-435.

[116] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, G. E. Hinton (1991). Adaptative mixtures of local experts. Neural Computation, 3(1):79-87.


29

[117] M. Alkoot, J. Kittler (1999). Experimental evaluation of expert fusion strategies. Pattern Recognition Letters, 20(11-13):1361-1369.

[118] Tax, M. Breukelen, R. Duin, J. Kittler (2000). Combining multiple classifiers by averaging or by multiplying? Pattern Recognition, 33(9):1475-1485.

[119] L. I. Kuncheva (2002). A theoretical study on six classifier fusion strategies. IEEE Trans. on Pattern Analysis and Machine Intelligence, 24(2):281-286.

[120] C. Y. Suen, C. Nadal, R. Legault, T. A. Mai, L. Lam (1992). Computer recognition of unconstrained handwritten numerals. Procs of IEEE, 80:1162-1180.

[121] L. Lam, C. Y. Suen (1995). Optimal combinations of pattern classifiers. Pattern Recognition Letters, 16(9):945-954.

[122] Gorski (1996). Practical combination of multiple classifiers. Progress in Handwritting Recognition, World Scientific, pp. 277-284.

[123] L. Breiman (1996). Bagging predictors. Machine Learning, 24(2):123-140, 1996. [124] T. K. Ho (1998). The random subspace method for constructing decision forests. IEEE

Trans. on Pattern Analysis and Machine Intelligence, 20(8):832-844. [125] D. W. Optiz (1999). Feature selection for ensembles. 16th Int. Conf. on Artificial

Intelligence, pages 379-384. [126] Valentini, T. G. Dietterich (2002). Bias-Variance Analysis and Ensembles of SVM. 3rd

Int. Workshop on Multiple Classifier Systems, pp. 222-231. [127] D. Partridge, W.B. Yates (1996). Engineering multiversion neural-net systems. Neural

Computation, 8(4):869–893. [128] Nishimura, M. Kobayashi, M. Maruyama, Y. Nakano (1999). Offline character

recognition using HMM by multiple directional feature extraction and voting with bagging algorithm. 5th ICDAR, pages 49–52.

[129] S. Gunter, H. Bunke (2003). Ensembles of classifiers for handwritten word recognition, Int. Journal on Document Analysis and Recognition, 5:224-232.

[130] L. S. Oliveira, R. Sabourin,, F. Bortolozzi, C. Y. Suen. (2003b). Feature Selection for Ensembles: A Hierarchical Multi-Objective Genetic Algorithm Approach, 7th Int. Conference on Document Analysis and Recognition, vol 2, pp. 676-680.

[131] M. Morita, L. S. Oliveira, R. Sabourin (2004). Unsupervised Feature Selection for Ensemble of Classifiers. IWFHR-9, pages 81-86.

[132] C-L Liu, K. Narukawa (2004b). Normalization Ensemble for Handwritten Character Recognition. 9th International Workshop on Frontiers of Handwriting Recognition, pp. 69-74

[133] Takahashi, T. Grifin (1993). Recogniton enhancement by linear tournament verification. 2nd Int. Conference on Document Analysis and Recognition, pp. 585-588.

[134] J. Zhou, Q. Gan, A. Krzyzak, C. Y. Suen. (2000) Recognition and verification of touching handwritten numerals. 7th International Workshop on Frontiers of Handwriting Recognition, pages 179-188.

[135] A.S. Britto Jr., R. Sabourin, F. Bortolozzi, C. Y. Suen (2002). A string lenght predictor to control the level building of HMMs for handwritten numeral recognition. 16th Int. Conference on Pattern Recognition, vol 4, pages 31-34.

[136] L. S. Oliveira, R. Sabourin, F. Bortolozzi, C. Y. Suen. (2003c). Impacts of verification on a numeral string recognition system. Pattern Recognition Letters, 24(7):1023-1031.

Document Analysis

30

[137] L. Vuurpijl, L. Schomaker, M. van Erp (2003). Architectures for detecting and solving conflicts: two-stage classification and support vector classifiers, Int. Journal on Document Aanalysis and Recognition, 5(4):213-223.

[138] Guyon, R. Haralick, J. Hull, I. Phillips (1997). Data sets for OCR and document image understanding research. Handbook of Character Recognition and Document Image Analysis, pages 780–799.

[139] A. W. Senior, A. J. Robinson (1998). An off-line cursive handwriting recognition system. IEEE Trans. on Patt. An. and Mach. Int., 20(3):309-321, March.

[140] U.-V. Marti, H. Bunke (2002). The IAM-database: an English sentence database for off-line handwriting recognition. Int. Journal on Document Aanalysis and Recognition, 5:39-46.

[141] M. Mori, A. Suzuki, A. Siho, S. Ohtsuka (2000). Generating new samples from handwritten numerals based on point correspondence. 7th International Workshop on Frontiers of Handwriting Recognition, pp. 281-290.

[142] Guyon (1996). Handwritten synthesis from handwritten glyphs. 5th International Workshop on Frontiers of Handwriting Recognition, pp. 309-312.

[143] M. Helmers, H. Bunke (2003). Generation and use of the synthetic training data in cursive handwriting recognition. 1st Iberian Conf. on Pattern Recognition and Image Analysis, pages 336-345.

[144] Varga, T., & Bunke, H. (2003). Generation of Synthetic Training Data for an HMM-based Handwriting Recognition System. In Proceedings of the 7th Int. Conference on Document Analysis and Recognition, Edinburgh, Scotland, pp. 618–622.

[145] Lorette (1999). Handwriting recognition or reading? What is the situation at the dawn of the 3rd millenium? Int. Journal on Document Analysis and Recognition, 2:2-12.

Recent Advances in Handwriting Recognition · Recent Advances in Handwriting Recognition 3 • The...

Documents

Transcript of Recent Advances in Handwriting Recognition · Recent Advances in Handwriting Recognition 3 • The...