Handwritten Japanese text recognition is still a big challenging task due to the large character set, diversity of writing styles, and multiple-touches between characters. In this paper, we propose a model of Attention Augmented Convolutional Recurrent Network (AACRN) for recognizing handwritten Japanese text lines. The AACRN model has three main parts: a convolutional feature extractor, a self-attention based encoder, and a CTC-decoder. The whole model can be trained end-to-end. In the experiment, we evaluate the performance of the AACRN model on the TUAT Kondate dataset and the Kuzushiji dataset. The results of the experiments show that the proposed model achieves higher performance than the state-of-the-art recognition accuracies on the test set of TUAT Kondate and the Kuzushiji dataset.

This paper presents an approach for matching various sizes of excavated wooden fragments based on siamese neural networks. We propose a siamese network composed of global average pooling and a fine-tuned Resnet encoder (GA-S-net). We also propose an elaborated siamese network by replacing the global average pooling with a spatial pyramid pooling layer and add a new dense absolute difference layer (SP-S-net). Samples of 37,760 fragments were prepared from 268 complete wooden tablets excavated from the Heijo-Kyo Palace ruins used during the Nara period in Japan. Both of the networks answer whether two fragments are from the same tablet or not. The result of both networks for the testing set is similar to AUC (Area under the curve) of ROC (Receiver Operating Characteristic) curve being around 90%. In AUC of large fragments, however, SP-S-net is better than GA-S-net with 97.1% versus 93.8%. These networks are rather new for dealing with various sizes of inputs for the matching problem.

Recently, the segmentation-free approach using Convolution Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for handwritten text recognition has been investigated by many research groups. Although good results are produced on some datasets, there remain some drawbacks. It is neither robust against the change of handwriting styles such as gaps between characters, stroke widths and shapes of characters nor stable for skewed, and curved text lines unless heavily trained by such patterns. In this paper, we propose a segmentation-based method for handwritten Japanese text recognition. The method employs a semantic segmentation model for precisely splitting text lines into single characters. The semantic segmentation model is based on an encoder-decoder architecture like U-Net, but we employ available techniques to improve the accuracy of pixel classification. They are a deeper encoder with ResNet 101, dilated convolutions and Spatial Pyramid Pooling. Subsequently, a CNN based OCR is used to recognize segmented character images. Finally, the recognized candidates are considered in a lattice diagram combined with a linguistic context. The result of experiments shows that the segmentation model increases the mean IoU significantly from 89.32% to 94.96% and the proposed approach is robust to the change of handwriting styles while a segmentation-free method is quite sensitive to them, resulting in the significant reduction of the recognition rate.

Discriminating ambiguous symbols in online handwritten mathematical expression is difficult without context. We propose a Bidirectional Recurrent Neural Network for segmenting and classifying online handwritten mathematical symbols. The context from forward and backward directions helps the classification model discriminate ambiguous symbols and improve recognition rates. The classification model is integrated into the Stochastic Context-Free Grammar recognition system for recognizing mathematical expressions. We show the effectiveness of the approach for improving symbol classification and segmentation on the CROHME 2016 dataset.

This paper presents an improvement in recognizing offline handwritten mathematical expressions (HMEs) by deep neural networks. We train it end-to-end using weakly supervised learning. The network has three parts: an encoder using a Convolutional Neural Network to encode high-level features from an input HME image; a decoder using gated recurrent units with attention to parse the high-level features and generate an output expression in the LaTeX format; and a symbol classifier to improve the localization and classification of the high-level features. Besides, we use the model ensemble method for the beam search process to average the probabilities from multiple models. For the dataset of the Competition on Recognition of Online Handwritten Mathematical Expressions (CROHME) 2014 and 2016, we have achieved an expression recognition rate of 53.65% and 51.96% correspondingly, which is 6 points better than without weakly supervised learning. Furthermore, when ensembling several models, the recognition rate of our method is increased to 55.68% for the CROHME 2014 testing set.

Collecting handwritten mathematical expressions (HMEs) generally requires a lot of time and effort for data preparation, data collection, annotation and so on. In this paper, we present a method for generating realistic HMEs in a wide variety of structures and styles from a LaTeX sequence or a MathML script using online isolated symbol patterns. Our method firstly positions all symbols in a symbol relation tree constructed from the input LaTeX or MathML script. Then, it places normalized online symbol patterns into the corresponding locations. A questionnaire-based experiment shows that the synthetic patterns are as clear and natural as the real patterns. Therefore, we can use the generated synthetic HME patterns for research on HME recognition and clustering.

This paper presents an attention-based convolutional sequence to sequence (ACseq2seq) model for recognizing an input image of multiple text lines from Japanese historical documents without explicit segmentation of lines. The recognition system has three main parts: a feature extractor using Convolutional Neural Network (CNN) to extract a feature sequence from an input image; an encoder employing bidirectional Long Short-Term Memory (BLSTM) to encode the feature sequence; and a decoder using a unidirectional LSTM with the attention mechanism to generate the final target text based on the attended pertinent features. We also introduce a residual LSTM network between the attention vector and softmax layer in the decoder. The system can be trained end-to-end by a standard cross-entropy loss function. In the experiment, we evaluate the performance of the ACseq2seq model on the anomalously deformed Kana datasets in the PRMU contest. The results of the experiments show that our proposed model achieves higher recognition accuracy than the state-of-the-art recognition methods on the anomalously deformed Kana datasets.

Despite of recent breakthroughs in the accuracy of single character recognition using the deeper convolution neural networks, one of the remaining problems is that OCRs almost fail to recognize character patterns when they are severely degraded, especially those of the historical documents. Another problem to recognize characters in historical documents is the lack of sufficient training patterns because of the heavy cost for annotation. This paper proposes a character attention generative adversarial network named CAGAN for restoring heavily degraded character patterns in historical documents so that OCRs improve their accuracy and even help archeologists to decode them. The network is based on the U-Net like architecture [1] with skip connections, and it is trained by the proposed loss function including the common adversarial loss (global loss) and the hierarchical character attentive loss (local loss). We made an experiment on 118 categories of most common Japanese Kanji characters, collected from severely damaged historical documents called Heijokyo mokkan written during the Nara period in Japan. The experiment shows that our method restores the shapes of characters and improves the recognition rate significantly, which is helpful for archeologists to decode damaged character patterns.

This paper presents an end-to-end model of Deep Convolutional Recurrent Network (DCRN) for recognizing offline handwritten Japanese text lines. The end-to-end DCRN model has three parts: a convolutional feature extractor using Deep Convolutional Neural Network (DCNN) to extract a feature sequence from a text line image; recurrent layers employing a Deep Bidirectional LSTM to predict pre-frame from the feature sequence; and a transcription layer using Connectionist Temporal Classification (CTC) to convert the pre-frame predictions into the label sequence. Since our end-to-end model requires a large data for training, we synthesize handwritten text line images from sentences in corpora and handwritten character patterns in the Nakayosi and Kuchibue database with elastic distortions. In the experiment, we evaluate the performance of the end-to-end model and the effectiveness of the synthetic data generation method on the test set of the TUAT Kondate database. The results of the experiments show that our end-to-end model achieves higher than the state-of-the-art recognition accuracy on the test set of TUAT Kondate with 96.35% and 98.05% character level recognition accuracies without and with the generated synthetic data, respectively.

This paper presents an attempt to recognize online isolated handwritten Japanese characters as well as text lines by Recurrent Neural Networks (RNNs). Although there are some successful studies on online Chinese handwriting recognition using RNNs, it is difficult to achieve high accuracy for Japanese due to many different types of characters such as kanji and kana. Moreover, training RNNs usually requires a large number of samples for each class of character. Hence, we apply five different transformation operations on the original samples to generate artificial samples with various deformations. For online handwritten Japanese text, we use the Kondate database, but it does not cover the whole Japanese character set. Thus, we generate random text lines using the sentences from corpora and isolated characters from the Nakayosi and Kuchibue databases. Besides, we employ a robust process with some preprocessing steps and the state-of-the-art online features to extract invariant features from handwritten patterns because the handwritten samples are merged from the different databases collected on different devices with various resolutions. For the recognition model, we have implemented the different Bidirectional Long Short-Term Memory Networks (BLSTM networks) for isolated character classification (sequence classification task) and handwritten text recognition (transcription task). The best model of the sequence classification task achieves an accuracy of 97.91% on Nakayosi and 97.74% on Kuchibue. The best model of transcription task performs a character recognition rate of 86.31% on Kondate and 83.15% on the generated text lines.

This paper presents the results of the VOHTR 2018 competition on Vietnamese Online Handwritten Text Recognition. The goal of this competition is to evaluate and compare recent online handwritten text recognition systems on Vietnamese online handwritten text which contains many delayed strokes caused by the diacritic marks. Besides, the general objective is to encourage the studies on Vietnamese online handwritten text recognition based on the large Vietnamese handwriting database collected from 200 writers. In this competition, we introduce three tasks consisting of word recognition (task 1), text-line recognition (task 2) and paragraph recognition (task 3) which are described in details. Subsequently, we describe the evaluation metrics and give comparative results of competitors along with the brief descriptions of the respective methods.

This paper proposes a method for speeding upparsing process for recognizing online handwritten mathematicalexpressions (OHME). We prune infeasible partitions in theparsing table to reduce the time for the parsing process. Lowscore partitions are candidates for pruning. Our method can beapplied for any parsing algorithms that use score functions. Inthis paper, we use a stroke order free system as a baseline system.The method is as follows. First, we analyze the scores ofpartitions in each row of the parsing table. Then, we determine athreshold for each row to prune low score partitions. Finally, weemploy these thresholds to prune low score partitions on thebaseline recognition system. The results of evaluations of ourmethod on the CROHME 2014 database show that therecognition process is speeded up by 3.46 times and 4.97 timeswhile recognition rate is reduced only 0.31 point and 0.71 point,respectively.

Motivated by recent successes in neural machine translation and image caption generation, we present an end-to-end system to recognize Online Handwritten Mathematical Expressions (OHMEs). Our system has three parts: a convolution neural network for feature extraction, a bidirectional LSTM for encoding extracted features, and an LSTM and an attention model for generating target LaTex. For recognizing complex structures, our system needs large data for training. We propose local and global distortion models for generating OHMEs from the CROHME database. We evaluate the end-to-end system on the CROHME database and the generated databases. The experiential results show that the end-to-end system achieves 28.09% and 35.19% recognition rates on CROHME without and with the generated data, respectively.

This paper proposes a modified X-Y cut method for reordering strokes of online handwritten mathematical expression (ME) in order to make stroke order free recognition. To deal with overlapping, which causes problems in the X-Y cut method, we determine vertically ordered strokes by detecting vertical symbols and its upper/lower MEs. An upper ME and a lower ME are treated as MEs which are reordered recursively. Unordered strokes on the left side of a vertical symbol are reordered as horizontally ordered strokes. The remaining strokes are reordered recursively. The horizontally ordered strokes are reordered from left to right and the vertically ordered strokes are reordered from top to bottom. The results of evaluations of the reordering method on the CROHME 2014 database show that our ME recognition system incorporating this method outperforms all other systems that use only CROHME 2014 for training while the processing time is kept to a practical level.

This paper presents our attempts to collect and analyze unconstrained Vietnamese online handwriting text patterns by pen-based computers. Totally, our database contains over 120,000 strokes from more than 140,000 characters, which is one of the largest Vietnamese online handwriting pattern databases currently. For building and analyzing our database, we made a collection tool, a line segmentation tool, and a delayed stroke detection tool. Moreover, we investigated some statistical information from personal information of writers. In order to solve the unconstrained handwriting recognition problem, we conducted experiments using Bidirectional Long Short-Term Memory (BLSTM) networks. BLSTM network is architecture of Recurrent Neural Network (RNN) and applied recently for many related problems. The performance of BLSTM network on our database is nearly 80% of accuracy even though this database contains many delayed strokes. In near future, we are going to avail our database for research purposes, as it would be the fundamental for the handwriting recognition research.

Parsing process is the most important process in recognition of online handwritten mathematical expressions. There are two basic approaches: stroke order dependent (SOD) and stroke order free (SOF) approaches. The SOD approach depends on stroke order while the SOF approach is free from stroke order. Although both approaches have shown high recognition rates in recently competitions, there are a few of works that analyze the complexities of parsing algorithms in the same experimental conditions. In this work, we have tested and analyzed recognition rate, recognition speed and memory space required by parsing algorithms on CROHME 2014. SOF is slightly superior to SOD in recognition rate, but SOD is faster in processing time and lower in memory space than SOF.

This paper presents a candidate lattice refinement method for online handwritten Japanese text recognition. In the integrated segmentation-recognition framework, we first over-segment a character string pattern into primitive segments at least at their true boundaries so that each primitive segment may compose a single character or a part of a character. Then a candidate lattice is constructed based on the primitive segments. We search within the candidate lattice to obtain the optimal path as recognition result. In striving for high recognition accuracy, however, the approach must generate many candidate lattice nodes, which ultimately increase the recognition time. To solve this problem, we refine the candidate lattice to eliminate unnecessary nodes before path search and text recognition. For the refinement, we evaluate all segmentation hypotheses by combining the probability of a character verifier using noncharacter samples, the class-independent unary and binary geometric context, as well as character segmentation. We retain N-best paths by beam search to reduce the complexity of the candidate lattice. Experiments on horizontal text lines extracted from the Kondate database show that the proposed method keeps recognition accuracy while reducing recognition time to half.

This paper presents Optical Handwritten Text Recognition (OHTR) for casually handwritten Japanese text. Distortions due to handwriting and the mixture of complex Chinese characters with simple phonetic and alphanumeric characters leave OHTR for handwritten Japanese text as still one of the hardest problems. This paper proposes a method for segmenting characters in handwritten text-pages, and a robust recognition model for over-segmentation. Stroke Width Transform (SWT), a new method of bridge separation at fork points and Voronoi diagrams with two improvements are proposed for character segmentation, and then a recognition model is employed with linguistic context and geometric context to recognize segmented characters. The results of experiments show that the proposed method improves the segmentation and recognition rates and each component in the method contributes to the improvement.

This paper presents a Finite State Machine (FSM) to reduce user's waiting time to get the recognition result after finishing writing in recognition of online handwritten English text. The lexicon is modeled by a FSM, and then determination and minimization are applied to reduce the number of states. The reduction of states in the FSM shortens the waiting time without degrading the recognition accuracy. Moreover, by merging incoming paths to each state, the recognition rate is improved. The N-best states decoding method also reduces the waiting time significantly with small degradation in recognition accuracy. Experiments on IAM-OnDB and IBM_UB_1 show the effectiveness of the method in both reducing waiting and improving recognition accuracy.

This paper presents a semi-incremental recognition method for online handwritten mathematical expressions (MEs). The method reduces the waiting time after an ME is written until the result of recognition is output. Our method has two main processes, one is to process the latest stroke, the other is to find and correct wrong recognitions in the strokes up to the latest stroke. In the first process, the segmentation, recognition and Cocke-Younger-Kasami (CYK) algorithm are only executed for the latest stroke. In the second process, all the previous segmentations are updated if they are significantly changed after the latest stroke is input, and then, all the symbols related to the updated segmentations will be updated with their recognition scores. These changes are reflected into the CYK table. In addition, the waiting time is further reduced by employing multi-thread processes. Experiments on our data set show the effectiveness of this semi-incremental method which not only has higher recognition rate than our previous pure-incremental method but also keeps the waiting time unnoticeable.

This paper describes an online handwritten cursive word recognition approach by combining segmentation-free and segmentation-based methods. To search the optimal segmentation and recognition path as the recognition result, we can attempt two methods: segmentation-free and segmentation-based, where we expand the search space using a character-synchronous beam search strategy. The probable search paths are evaluated by integrating character recognition scores with geometric characteristics of the character patterns in a Conditional Random Field (CRF) model. We make a comparison between online handwritten cursive word recognition using segmentation-free method and that using segmentation-based method, and then attempt combining the two methods to improve performance. Our methods restrict the search paths from the trie lexicon of words and preceding paths during path search. We show this comparison on a publicly available dataset (IAM-OnDB).

This paper presents a system for recognizing online handwritten mathematical expressions (MEs) and improvement of structure analysis. We represent MEs in Context Free Grammars (CFGs) and employ the Cocke-Younger-Kasami (CYK) algorithm to parse 2D structure of on-line handwritten MEs and select the best interpretation in terms of symbol segmentation, recognition and structure analysis. We propose a method to learn structural relations from training patterns without any heuristic decisions by using two SVM models. We employ stroke order to reduce the complexity of the parsing algorithm. Moreover, we revise structure analysis. Even though CFG does not resolve ambiguities in some cases, our method still gives users a list of candidates that contain expecting result. We evaluate our method in the CROHME 2013 database and demonstrate the improvement of our system in recognition rate as well as processing time.

We present our recent model of a diagram recognition engine. It extends our previous work which approaches the structural recognition as an optimization problem of choosing the best subset of symbol candidates. The main improvement is the integration of our own text separator into the pipeline to deal with text blocks occurring in diagrams. Second improvement is splitting the symbol candidates detection into two stages: uniform symbols detection and arrows detection. Text recognition is left for post processing when the diagram structure is already known. Training and testing of the engine was done on a freely available benchmark database of flowcharts. We correctly segmented and recognized 93.0% of the symbols having 55.1% of the diagrams recognized without any error. Considering correct stroke labeling, we achieved the precision of 95.7%. This result is superior to the state-of-the-art method with the precision of 92.4%. Additionally, we demonstrate the generality of the proposed method by adapting the system to finite automata domain and evaluating it on own database of such diagrams.

This paper describes significant improvement in On-line handwritten Japanese text recognition that is free from line direction and character orientation constraints. The original system [1, 2] separates freely written text into text line elements, estimates and normalizes character orientation and line direction. Then, it hypothetically segments each text line element into primitive segments, constructs a segmentation-recognition candidate lattice and evaluates the likelihood of candidate segmentation-recognition paths by combining the scores of character recognition, geometric features, as well as linguistic context. In this scheme, we have updated the over-segmentation for each text line element and applied a robust context integration model to recognize each text line element. Experimental results on text from the HANDS-Kondate_t_bf-2001-11 database demonstrate large improvement in the character recognition rate compared with the previous system [1, 2].

In the paper, we compare the methods based on Markov Random fields (MRF) and Conditional Random fields (CRF) for separating text and non-text ink strokes in online handwritten Japanese documents. This paper validates the effect of context information in neighbor strokes based on graphical models of MRF and CRF. The task of separating text and non-text ink strokes in ink documents denotes classifying ink strokes into two classes (text and non-text). For classification, Support Vector Machine (SVM) classifiers are trained on the set of ink strokes. After converting the SVM's outputs to likelihood probabilities, they are assigned to the likelihood clique potentials of MRF and the feature functions of CRF. The classification based on MRF or CRF is considered as a labeling problem, which can be solved using a labeling algorithm. The experiments on Japanese ink documents in the Kondate database shows that the proposed method based on CRF achieves a classification rate of 98.02% while the method based on MRF produces the classification rate of 97.86%.

This paper describes a database of on-line handwritten patterns mixed of text, figures, tables, maps, diagrams and so on. Now, pen-based and touch-based interfaces are spreading into people and their surfaces are getting large. People can write and draw mixed objects without paying attention on the difference of objects or the mode change. Moreover, they may write text in any direction in combination with non-text objects on large surfaces. This is clearly one of the largest advantages of pen or touch interfaces but poses a challenging problem of object classification and recognition. The proposed database is made and now being enlarged to study such subjects more extensively. So far, 100 Japanese writers, approximately 25 English and 45 Thai writers have participated. The database stores on-line handwritten (digital ink) patterns with ground-truth tags in InkML.

This paper presents a semi-incremental recognition method for online handwritten English text. We employ local processing strategy and focus on a recent sequence of strokes defined as "scope". For the latest scope, we build and update a segmentation and recognition candidate lattice and advance the best-path search incrementally. We utilize the result of the best-path search in the previous scope to exclude unnecessary segmentation candidates. This reduces the number of candidate word recognition with the result of reduced processing time. We also reuse the segmentation and recognition candidate lattice in the previous scope for the latest scope. Moreover, triggering recognition processes every few strokes save CPU time. Experiment made on IAM-OnDB database shows the effectiveness of the proposed method not only in reduced processing time and waiting time, but also in recognition accuracy.

In this paper, we propose a novel method for text/non-text classification in online handwritten document based on Recurrent Neural Network (RNN) and its improved version, Long Short-Term Memory (LSTM) network. The task of classifying strokes in a digital ink document into two classes (text and non-text) can be seen as a sequence labelling task. The bidirectional architecture is used in these networks to access to the complete global context of the sequence being classified. Moreover, a simple but effective model is adopted for the temporal local context of adjacent strokes. By integrating local context and global context, the classification accuracy is improved. In our experiments on the Japanese ink documents (Kondate database), the proposed method achieves a classification rate of 98.75%, which is significantly higher than the 96.61% in the previous work. Similarly, on the English ink documents (IAMonDo database), it produces a classification rate of 97.68%, which is also higher than other results reported in the literature.

This paper presents a semi-incremental recognition method for online Japanese handwritten text recognition, which is used for busy recognition interface (recognition while writing) and lazy recognition interface (recognition after writing) without large waiting time. We employ local processing strategy and focus on a recent sequence of strokes defined as "scope". For the latest scope, we build and update a segmentation and recognition candidate lattice and advance the best-path search incrementally. We utilize the result of the best-path search in the previous scope to exclude unnecessary segmentation candidates. This reduces the number of candidate character recognition with the result of reduced processing time. We also reuse the segmentation and recognition candidate lattice in the previous scope for the latest scope. Moreover, triggering recognition processes every few strokes save CPU time. Experiment made on TUAT-Kondate database shows the effectiveness of the proposed method not only in reduced processing time and waiting time, but also in recognition accuracy.

We propose a method which performs an isotropic morphological dilation via implicit smoothing for the purpose of restoring the degraded character shapes of binarized images. Exploiting the idea of geodesic morphology that the binary image and its distance transformed image are interconvertible, we apply a smoothing method not to the binary image but to the distance transformed image, and then reconvert it by binarization. This allows us to apply conventional smoothing methods for continuous intensity, i.e., gray scale, images to the discrete intensity, i.e., binary, image implicitly. For instance, by using an isotropic diffusion together with geodesic dilation, an isotropic dilation along the stroke direction is obtained and brings better results.

We propose a segmentation based online word recognition approach which uses a Conditional Random Field (CRF) driven beam search strategy. An efficient trie-lexicon directed, breadth-first beam search algorithm is employed in a combined segmentation-and-recognition framework to accomplish real-time recognition of online handwritten cursive English words. This framework is developed by building a candidate lattice of primitive segments obtained through over segmentation of the word pattern. The search space for the lattice is expanded by synchronously matching the lattice nodes to likely character patterns from a trie-dictionary constructed out of the target lexicon. The probable paths are evaluated by integrating character recognition scores with physical and spatial characteristics of the handwritten segments in a CRF (conditional random field) model and a beam search strategy is used to prune the set of likely paths. This approach has been benchmarked on the new IBM_UB_1 dataset as well as on the UNIPEN dataset for comparison.

This paper describes an online handwritten English cursive word recognition method using a segmentation-free Markov random field (MRF) model in combination with an offline recognition method which uses pseudo 2D bi-moment normalization (P2DBMN) and modified quadratic discriminant function (MQDF). It extracts feature points along the pen-tip trace from pen-down to pen-up and uses the feature point coordinates as unary features and the differences in coordinates between the neighboring feature points as binary features. Each character is modeled as a MRF and word MRFs are constructed by concatenating character MRFs according to a trie lexicon of words during recognition. Our method expands the search space using a character-synchronous beam search strategy to search the segmentation and recognition paths. This method restricts the search paths from the trie lexicon of words and preceding paths, as well as the lengths of feature points during path search. We also combine it with a P2DBMN-MQDF recognizer that is widely used for Chinese and Japanese character recognition.

This paper proposes a method for handwritten Chinese/Japanese text (character string) recognition based on semi-Markov conditional random fields (semi-CRFs). The high-order semi-CRF model is defined on a lattice containing all possible segmentation-recognition hypotheses of a string to elegantly fuse the scores of candidate character recognition and the compatibilities of geometric and linguistic contexts by representing them in the feature functions. Based on given models of character recognition and compatibilities, the fusion parameters are optimized by minimizing the negative log-likelihood loss with a margin term on a training string sample set. A forward-backward lattice pruning algorithm is proposed to reduce the computation in training when trigram language models are used, and beam search techniques are investigated to accelerate the decoding speed. We evaluate the performance of the proposed method on unconstrained online handwritten text lines of three databases. On the test sets of databases CASIA-OLHWDB (Chinese) and TUAT Kondate (Japanese), the character level correct rates are 95.20 and 95.44 percent, and the accurate rates are 94.54 and 94.55 percent, respectively. On the test set (online handwritten texts) of ICDAR 2011 Chinese handwriting recognition competition, the proposed method outperforms the best system in competition.

We have many historical documents written in over 1,000 years ago. Shape features of character patterns on the documents are unstable or missing because most of the documents have been stained and degraded deeply. Digital archives of the documents with accurate character pattern retrieval methods are helpful for archaeologists and historians. In this paper, we propose a similarity evaluation method for character patterns with missing shape parts. It collaboratively works with non-linear normalization for such patterns, and modifies the templates for each trial of the retrieval efficiently. In the experiences using 4,911 Kanji (Chinese origin) character patterns from the Japanese historical documents called mokkans, the method shows improvements of the retrieval accuracy. Also, we present a simple implementation of gradient feature extraction to compare the chain code feature with the gradient feature in the retrieval. As the result, the gradient feature works better than the chain code feature.

This paper presents the effect of text/non-text classification for ink search which employs string recognition. Pen or touch interfaces provides the benefit that users can write text and draw figures without changing the device or mode, but line drawings are troublesome for ink search. We propose the insertion of text/non-text classification before ink search and show its effect. For ink search, we employ our own engine to search keywords in the candidate lattice prepared by on-line handwritten Japanese text recognition, since this method produces higher search rate for Japanese text in digital ink than word spotting without ink recognition.

In this paper, we present methods of segmenting Nom historical documents and clustering character patterns to build a Nom character pattern database. Nom is an ideographic script to represent Vietnamese, used from the 10th century to 20th century. However, this heritage is nearly lost. In order to preserve the wisdom and knowledge expressed in Nom, recognition and digitalization are indispensable. Because there is no OCR for Nom yet, we have to start from collecting patterns. We have employed a projection profile based method for segmenting hundreds of pages into individual characters. Then, we have implemented a combination of Chinese OCR-based clustering and K-means clustering to group characters into categories. The experiment shows that the proposed system can help collecting the characters patterns effectively. Moreover, it has revealed that there are many character classes lost or uncategorized so far.

This paper presents a character recognition based approach to search for a keyword in on-line handwritten Japanese text. It employs an on-line character recognizer or an off-line recognizer, produces recognition candidates and search for a keyword in the lattice of the candidates. This paper also presents a feature matching based approach employing on-line features or off-line features. We compare the above two approaches and conclude that the character recognition based approach yields superior performance compared to the feature-matching-based approach.

This paper describes a method for building a compact on-line Markov random field (MRF) recognizer for large handwritten Japanese character set using structured dictionary representation and vector quantization (VQ) technique. The method splits character patterns into radicals, whose models by MRF are shared by different characters such that a character model is constructed from the constituent radical models. Many distinct radicals are shared by many characters with the result that the storage space of model dictionary can be saved. Moreover, in order to further compress the parameters, we employ VQ technique to cluster parameter sets of the mean vectors and covariance matrixes for MRF unary features and binary features as well as the transition probabilities of each state into groups. By sharing a common parameter set for each group, the dictionary of the MRF recognizer can be greatly compressed without recognition accuracy loss.

This paper describes effects of a large amount of artificial patterns to train an on-line handwritten Japanese character recognizer. In general, as more learning patterns employed for training pattern recognition systems, as higher recognition rate is obtained. In reality, however, the existing pattern samples are not enough, especially for languages of a large character set. Therefore, for on-line handwritten Japanese character recognition, we construct six linear distortion models and combine them with a nonlinear distortion model to generate a large amount of artificial patterns. We apply the method for the TUAT Nakayosi database and train a recognizer while evaluate the effects for the TUAT Kuchibue database with the remarkable effects of improving recognition accuracy.

This paper describes an unconstrained on-line handwritten Japanese text retrieval system from character recognition candidates. The system is based on a discriminative model which integrates the scores of character recognition, segmentation and geometric context in search and retrieval, and the parameters are trained by supervised learning. Experiments on TUAT Kuchibue database show that the proposed method can effectively improve the system performance. When the search method with the optimal threshold retrieves for a keyword consisting of two, three or four characters, its f-measure is 0.720, 0.868 or 0.923, respectively.

In this paper, we describe a method of retrieving on-line handwritten text based on directional feature matching. Although text search into the character recognition candidate lattice has been elaborated, the character recognition based approach does not support languages which are not assumed. The proposed method is liberated from this constraint. It first hypothetically segments on-line handwritten text into character pattern blocks and prepares the object text patterns by combining the character pattern blocks. On the other hand, it employs handwritten text as a query pattern or prepares a query pattern by combining character ink patterns from query character codes. Then, it extracts directional features from the object text patterns and the query pattern, and the dimensionalities of those features are further reduced by Fisher linear discriminate analysis (FDA). Finally, the similarity is measured between the object text patterns and the query pattern by block-shift matching. This paper discusses the retrieval performance in comparison with our previous character recognition based method.

In offline handwritten character recognition, the nonlinear normalization (NLN) method based on line density equalization has been proven very effective. This paper shows the effects on online handwritten Japanese character recognition. We apply the nonlinear normalization based on line density equalization to online character patterns. Since the curve-fitting-based normalization methods and their pseudo 2D extensions yields superior performance on offline patterns, we also combine these methods with the way using line density projection. We have compared the methods using trajectory-based projection with ones using line density projection. As a result, line density-based methods yield superior accuracy and a competitive time-complexity.

This paper describes the design and implementation of a recognition server for online handwritten Japanese text. Currently, fast and accurate recognition of online handwritten characters requires a high-performance CPU and a large memory space. However, PDA and other small portable devices lack the qualification. On the other hand, data transmission speed is getting remarkably higher. To solve the CPU and memory constraint on a client and avail the speed-up of transmission, we employ a network environment and build an a recognition server. This paper descrives network data flow transmission, client/server architecture and presents performance.

This paper presents a conditional random field (CRF) model for aligning online handwritten Chinese/Japanese text lines (character strings) with the corresponding transcripts. The CRF model is defined on a lattice which contains all possible segmentation hypotheses. The feature functions characterize the shape and context dependences of characters, including the scores of character recognition and the geometric compatibilities between characters. The combining parameters are optimized by energy minimization. Experimental results on two online databases: CASIA-OLHWDB and TUAT Kondate demonstrate the effectiveness of the proposed method.

This paper describes effective object function design for combining on-line and off-line character recognizers for on-line handwritten Japanese text recognition. We combine on-line and off-line recognizers using a linear or nonlinear function with weighting parameters optimized by the MCE criterion. We apply a k-means method to cluster the parameters of all character categories into groups so that the categories belonging to the same group have the same weight parameters. Moreover, we apply a genetic algorithm to estimate super parameters such as the number of clusters, initial learning rate and maximum learning times as well as the sigmoid function parameter for MCE optimization. Experimental results on horizontal text lines extracted from the TUAT Kondate database demonstrate the superiority of our method.

This paper describes a Markov random field (MRF) model with weighting parameters optimized by conditional random field (CRF) for on-line recognition of handwritten Japanese characters. It also presents updated evaluation using a large testing set. The model extracts feature points along the pen-tip trace from pen-down to pen-up and sets each feature point from an input pattern as a site and each state from a character class as a label. It employs the coordinates of feature points as unary features and the differences in coordinates between the neighboring feature points as binary features. The weighting parameters are estimated by CRF or the minimum classification error (MCE) method. In experiments using the TUAT Kuchibue database, the method achieved a character recognition rate of 92.77%, which is higher than the previous model's rate, and the method of estimating the weighting parameters using CRF was more accurate than using MCE.

This paper describes a method for constructing the most efficient and robust coarse classifier from a large number of basic recognizers which are obtained by different parameters of feature extraction, different discriminant methods or functions, and so on. The architecture of the coarse classification is a sequential cascade of basic recognizers and reduces the candidates after each basic recognizer. Genetic algorithm determines the best cascade with the best speed and highest performance. The method is applied for on-line handwritten Japanese characters recognition. We produced 201 basic recognizers of MQDF, 21 basic recognizers of Euclidian distance and 21 basic recognizers of the LSS method by changing parameters. From these basic recognizers we have obtained a rather simple 2 stages cascade with the result that the whole recognition time was reduced to 24.5% while keeping classification and recognition rates.

This paper describes a lexicon-driven approach to on-line handwritten Japanese disease name recognition using a time-synchronous method. A trie lexicon is constructed from a disease name database containing 21,713 disease name phrases. It expands the search space using a time-synchronous method and applies the beam search strategy to search into a segmentation candidate lattice constructed based on primitive segments. This method restricts the character categories for recognizing each character candidate pattern from the trie lexicon of disease names and preceding paths during path search in the segmentation candidate lattice, and selects an optimal disease name from the disease name database as recognition result. The experimental results demonstrate the effectiveness of our proposed method, which improves character recognition rate from 94.56% to 99.97% compared with a general-purpose Japanese text recognizer and speeds up recognition time as 4.3 times faster as the general recognizer.

In this study, we investigate several methods for enhancing scanned mokkan images to aid archeologists and historians in the interpretation of mokkans. Mokkans are wooden tablets with handwritten characters used in 8th century Japan. Due to damages and natural deterioration, the interpretation of the text on mokkans is difficult even for archeologists and historians. The automatic interpretation of mokkan text is unrealistic due to drastic deterioration. Our aim is to realize an image processing software system for archeologists, specific to the processing of mokkans. This paper evaluates noise reduction and contrast enhancement techniques for inclusion in the system. Several operations such as thresholding and Gaussian blurring are shown to be effective, when used in conjunction with contrast enhancement in the HSV space. We also identified several techniques that allow convenient selection of character strokes in of mokkan images. The selection can then be used as input for character recognition.

To reduce the classification errors of online handwritten Japanese character recognition, we propose a method for confusing characters discrimination with little additional costs. After building confusing sets by cross validation using a baseline quadratic classifier, a logistic regression (LR) classifier is trained to discriminate the characters in each set. The LR classifier uses subspace features selected from existing vectors of the baseline classifier, thus has no extra parameters except the weights, which consumes a small storage space compared to the baseline classifier. In experiments on the TUAT HANDS databases with the modified quadratic discriminant function (MQDF) as baseline classifier, the proposed method has largely reduced the confusion caused by non-Kanji characters.

This system development enables ubiquitous peer review and reuse of reviewer comments to assess teaching/learning in higher education. The purpose of this project is two-fold: (1) to develop applications for recording and storing reviewer annotations on streaming class lectures as time sequence data of pen-tip coordinates, and (2) to identify key principles and criteria from annotated video data to assess and evaluate the quality of teaching and learning (e-teaching portfolio). The evaluation studies were conducted to gain a broad understanding of how reviewers identify and record educational events effectively and appropriately during ldquolesson study (Jyugyo kenkyu)rdquo when using this system. The collection and analysis of reviewers' annotations indicated that this system is capable of reusing collected comments in order to suggest weak and strong points in class lectures from different reviewers' perspectives.

This paper presents a revised method for keyword search from handwritten digital ink in comparison with the previous system. We adopt a search method using noise reduction. Experiments on digital ink databases show that the revised method typically improves the systempsilas overall accuracy (f-measure) from 0.653 to 0.891.

This paper describes an online handwritten Japanese character string recognition system based on conditional random fields, which integrates the information of character recognition, linguistic context and geometric context in a principled framework, and can effectively overcome the variable length of candidate segmentation. For geometric context, we employ both unary and binary feature functions, as well as the ones relevant and irrelevant to character classes. Experimental results show that the CRF based method outperforms the method with normalized path evaluation criterion, and the geometric context benefits the performance significantly.

This paper describes a method of on-line handwritten Japanese text recognition by improved path evaluation. Based on a theoretical ground, the method evaluates the likelihood of candidate segmentation paths by combining scores of character pattern size, inner gap, character recognition, single and pair character position, candidate segmentation point and linguistic context, with the weight parameters optimized by a genetic algorithm. The path score is insensitive to the number of candidate patterns and the optimal path can be found by Viterbi search. Experimental results demonstrate the superiority of the proposed method.

High-definition video applications, such as digital TV and digital video cameras, require high processing performance for high-quality visual images in addition to a complex video CODEC. Pre-/postprocessing to improve video quality is becoming much more important because requirements for pre-/postprocessing vary among applications and processing algorithms have not been stabilized. Therefore, a new processor architecture that has a highly parallel datapath is needed. In this paper, we introduce a VLIW vector media coprocessor, “vector coprocessor (VCP),” that includes three asymmetric execution pipelines with cascaded SIMD ALUs. To improve performance efficiency, we reduce the area ratio of the control circuit while increasing the ratio of the arithmetic circuit. The total gate count of VCP is 1268 kgates and its maximum operating frequency is 300 MHz at 90-nm CMOS process. Some of the processing kernels in an adaptive prefilter that is applied to preprocessing for video encoding are evaluated. In the case of the edgeness and the sum of absolute differences, the performance is 183 giga operations per second. VCP offers enough performance for HD video processing and good cost-performance while all processing pipeline units operate effectively.

This paper presents a design and an implementation of document image retrieval to support reading mokkans. A mokkan is a wooden tablet with text written by a brush in India ink. Despite the archaeological and historical value of the mokkans excavated from ancient ruins, many of the mokkans have not been decoded yet due to the lost or too much damaged character patterns on them. Character recognition for damaged patterns is useful to decode such mokkans. Furthermore, if the recognition results show not only the character codes but also the images of the character patterns and the whole mokkans, the recognition becomes useful document retrieval to complement the lost or unreadable part of the mokkans. In the implementation, we built a public database of historical mokkans with their photographs and a character recognition module working on our support system to search the database. The evaluation by archaeologists is in progress.

This paper presents development of a Web-based visual simulator with communication support and its application to computer architecture education. The simulator written in Java has equipped visualization and communication support for real education. The former is designed to show how a computer works graphically and helps learners to understand basic computer architecture visually. The latter is realized with a built-in e-mail function. It allows learners to send question e-mails with a snapshot of simulator's execution to instructors. They receive such question e-mails and then redo to verify the execution of learners' simulators. With these supporting equipments, learners and their instructors can share suitable information between their simulators.

In this paper we present a scoring tool for electronic paper exams. Digital ink written by students in exams is captured and then sent to the Paper Architecture server. There the ink is segmented and processed. We have taken a hybrid approach for scoring. Selection questions are automatically recognized and scored. For free text answering we have developed an interface that allows teachers to correct them interactively. The interface uses question reorganization and empty questions marking to reduce the marking time. An evaluation experiment comparing the exam marking by hand and by using the tool showed considerable reduction of the scoring time when using the tool.

This paper presents a visualization tool to improve the performance of a classifier based on the hidden Markov Model. A specific recognition system for which the visual- ization tool is designed is an on-line handwritten Japanese character recognition system. The recognition system was built from already estimated parameter values, which leads to some difficulties when trying to adjust the system. To tackle this problem we describe how visual information can be helpful to interpret the results and how it can be used to build a set of viewers for helping the tuning task. These viewers were used to examine the data structure and inter- nal procedures of the recognition engine allowing to detect and correct errors in the first implementation. We conclude the paper comparing the two implemented versions of the classifier by showing the increase we achieved in recogni- tion accuracy.

This paper describes a support system for archeologists to read "mokkan". A mokkan is a wooden tablet on which text was written by a brush. Many mokkans used in Nara period (from AD. 710 to 794) are being excavated from Heijyo-kyo, Japan (the ancient court in the Nara period). The support system is for archeologists who read mokkans that have been stained, damaged and degraded under the soil. Such mokkans are hard to read even for expert readers. However, the binarization functions of the system extract ink from the image of the mokkans and the character recognition function outputs candidates even for degraded or partially missing character patterns. We made also a graphical user interface to invoke the above functions, provide experts with suggestions and stimulate their inference. Archeologists in the experiment for evaluation enthusiastically accepted the system.

This paper describes a segmentation method of online handwritten Japanese text of arbitrary line direction by a neural network to improve text recognition performance. This method extracts multidimensional features from strokes of handwritten text and input them into a neural network to preliminarily determine segmentation points. Then, it modifies segmentation candidates using some spatial features. We compare the method with the previous method and that by Fisher's linear discriminant using the database HANDS-Kondate/spl I.bar/t/spl I.bar/bf-2001-11. This paper also shows how to generate character segmentation candidates in order to achieve high discrimination rate by investigating the relationship between recall, precision and the f measure.

This paper describes a method for separating online handwritten patterns into Japanese text, figures and mathematical formulas. Today, Tablet PCs and electronic whiteboards provide much larger writing area for pen interfaces unlike PDAs (Personal Digital Assistants), through which users can easily input text, write mathematical formulas and draw figures on the screen. The fact that these objects can be written by a single pen (marker) without switching the device, mode, software or whatever else, and without any writing restrictions such as grids or boxes is one of the most important benefits of the pen interfaces. However, the task of segmenting these objects is challenging. To address this issue, we have applied a probabilistic model employing stroke features, stroke crossings and stroke densities. Further, we partially apply the approach of segmentation by recognition. Although the current recognizer for formulas is not a true recognizer, we have achieved about 81% correct segmentation for all the strokes when applied to our newly prepared database of mixed patterns. This method has been compared with a neural-network. The results show that our method is generally better but less effective in distinguishing figures from other components.

Online handwriting recognition is gaining renewed interest owing to the increase of pen computing applications and new pen input devices. The recognition of Chinese characters is different from western handwriting recognition and poses a special challenge. To provide an overview of the technical status and inspire future research, this paper reviews the advances in online Chinese character recognition (OLCCR), with emphasis on the research works from the 1990s. Compared to the research in the 1980s, the research efforts in the 1990s aimed to further relax the constraints of handwriting, namely, the adherence to standard stroke orders and stroke numbers and the restriction of recognition to isolated characters only. The target of recognition has shifted from regular script to fluent script in order to better meet the requirements of practical applications. The research works are reviewed in terms of pattern representation, character classification, learning/adaptation, and contextual processing. We compare important results and discuss possible directions of future research.

We describe here information encoding and decoding methods applied to dot texture for active forms. We employ dot texture made of tiny dots and looking like gray color to print various forms. This facilitates the separation of handwriting from its input frame even under monochrome printing/reading environments. It also makes the forms determine how to process filled-in handwriting according to the information embedded in the dot texture. The embedded information results in an improved recognition rate of handwriting, and allows the form processing to be directed by the form itself rather than by the form reading machine. Thus, the form-reading machine becomes a general-purpose machine allowing different forms inputted into it to be processed differently as specified by each form. We compare various dot shapes and information encoding/decoding methods for those shapes. Then, we present how to locate input frames, separate handwriting from input frames and segment handwriting into characters. We also present preliminary evaluation of the described methods.

Despite all the attempts of modern information technologyto supplant paper by sophisticated electronic devices,paper is still omnipresent in virtually any office environment.We have learned over the last decade that paper isan indispensable part of our daily work and that even thelatest computers are not able to mimic its functionality. Today,paper and computer are coexisting side by side withoutinteracting with each other, both having their specificadvantages. This paper presents a concept for bridging thegap between computer and paper by combining the advantagesof both electronic and paper world. The implementationdescribed in this paper augments standard paper withfeatures of modern information technology. Moreover, it allowsinterfacing computers using pen and paper.

This paper describes an on-line handwritten Japanese text recognition method that is liberated from constraints on writing direction (line direction) and character orientation. This method estimates the line direction and character orientation using the time sequence information of pen-tip coordinates and employs writing-box-free recognition with context processing combined. The method can cope with a mixture of vertical, horizontal and skewed lines with arbitrary character orientations. It is expected useful for tablet PCs, interactive electronic whiteboards and so on.

This paper describes prototyping of a processingsystem employing dot texture for printing input frames ofthe form. The dot texture is the texture composed of smalldots. It eases the separation of handwritings from theinput frames even under monochrome printing/readingenvironments and makes the system to process thehandwritings according to the information embedded inthe dot texture of the frames. The embedded informationin the dictates how to process the so that wecall the active form being opposite to the passiveform processed by the program stored in a documentreader. This method can also be used to embed otherinformation such as attribute of handwriting and so on.This paper presents the design, prototyping and somepreliminary evaluation.

This paper proposes a new approach to accelerating speed and increasing the recognition rate of an off-line recognizer employed for on-line handwriting recognition of Japanese characters. All training patterns are divided according their stroke number to several groups and one single recognizer is dedicated for each group of patterns. Since a number of categories for a single recognizer is smaller, the speed and accuracy improves. First, we make the model of a recognizer and show that our method can theoretically accelerate its recognition speed to 45 % of the original time. Then, we employ the method to a practically used off-line recognizer with the result that the recognition rate is increased from 90.73 % to 91.60 % and the recognition time is reduced to only 49.73 % of the original one. Another benefit of our new approach is high scalability so that the recognizer can be optimized for speed and size or for the best accuracy. 1.

This paper proposes a method to accelerate character recognition of a large character set by employing pivots into the search space. We divide the feature space of character categories into smaller clusters and derive the centroid of each cluster as a pivot. Given an input pattern, it is compared with all the pivots and only a limited number of clusters whose pivots have higher similarities (or smaller distances) to the input pattern are searched for with the result that we can accelerate the recognition speed. This is based on the assumption that the search space is a distance space. The method has been applied to pre-classification of a practical off-line Japanese character recognizer with the result that the pre-classification time is reduced to 61 % while keeping its pre-classification recognition rate up to 40 candidates as the same as the original 99.6% and the total recognition time is reduced to 70% of the original time without sacrificing the recognition rate at all. If we sacrifice the pre-classification rate from 99.6% to 97.7%, then its time is reduced to 35% and the total recognition time is reduced to 51.5% with recognition rate as 96.3% from 98.3%.

This paper presents the UP-Kuchibue and UP-Nakayosi databases containing on-line handwritten Japanese characters. These databases are the international versions of two databases, Kuchibue and Nakayosi, collected in the Nakagawa Laboratory at the University of Agriculture & Technology in Tokyo. They contain more than 3 million characters written by 283 Japanese writers. UP-Kuchibue and UP-Nakayosi are stored in the common Unipen format. Unipen is a western plain ASCII format which allows easy access for international researchers and facilitates international benchmarks.

This paper describes our second collection of online handwritten character patterns and their analysis. 163 writers presented about 10,000 character patterns, covering 4,438 categories mainly in the context of sentences. Together with our first collection, the Kuchibue database containing 12,000 patterns from 120 writers, we have now collected about 3 million patterns. For this second collection of online patterns, named Nakayosi, we analyzed stroke number and order variations.

The availability of a large sample database is very important to design high accuracy classifiers for handwritten character recognition. Collecting image samples from human writers and practical documents is expensive particularly for large character sets, like with East-Asia-languages. We can therefore take advantage of existing online databases to generate additional off-line images. This paper proposes a method to generate realistic character images from online patterns. From the pen trajectory of an online pattern, the proposed method can generate numerous images of various stroke shapes using three painting modes: constant line mode, proportional mode and calligraphic mode. Particularly, the calligraphic mode combines the pen trajectory (representing the writing style of one concrete writer) with real stroke images (also representing individual writing style of a concrete writer) to generate character images that look as if they were produced with a brush or pen by human hand.

An online writing-box free method for recognizing handwritten Japanese text is proposed. This method is achieved by the following procedure. First, the average character size of input handwritten text is estimated. Second, candidates for character segmentation are detected using geometric features between two adjacent strokes. Finally, a search is performed by dynamic programming for the string that maximizes evaluation score (acceptability as Japanese text). The evaluation score reflects the likelihood of character segmentation, recognition, context and the size of each character. The size is a newly introduced factor since alphabets, numerals, symbols, Japanese phonetic characters, simple Chinese characters and compound Chinese characters are all written in different sizes even in a single line of text. Experimental results show that the incorporation of the character size likelihood increases the character recognition rate for Japanese text.

The smart VISIOn chip has a large potential for application in general purpose high speed image processing systems. In order to fabricate smart vision chips including photo detector compactly, we have proposed the application of three dimensional LSI technology for smart vision chips. Three dimensional technology has great potential to realize new neuromorphic systems inspired by not only the biological function but also the biological structure. In this paper, we describe our three dimensional LSI technology for neuromorphic circuits and the design of smart vision chips .

This paper proposes a precise candidate selection method for large character set recognition by confidence evaluation of distance-based classifiers. The proposed method is applicable to a wide variety of distance metrics and experiments on Euclidean distance and city block distance have achieved promising results. By confidence evaluation, the distribution of distances is analyzed to derive the probabilities of classes in two steps: output probability evaluation and input probability inference. Using the input probabilities as confidences, several selection rules have been tested and the rule that selects the classes with high confidence ratio to the first rank class produced best results. The experiments were implemented on the ETL9B database and the results show that the proposed method selects about one-fourth as many candidates with accuracy preserved compared to the conventional method that selects a fixed number of candidates.

Artificial neural networks have been widely used in handwritten numeral recognition (HNR) with some success. This paper presents some new results of HNR using MLP (multilayer perceptron) and RBF (radial basis function) neural networks. By using discriminative training, which aims to minimize the empirical classification error on a training data set, the recognition accuracy is considerably improved. The performance of the RBF net is comparable to that of the MLP in terms of the forced recognition rate, and even better than the MLP in terms of the rejection-error tradeoff. The experiments were implemented on the CENPARMI database, and very high recognition rates have been obtained.

This paper reviews some prototype learning algorithms for nearest neighbor (NN) classifier design land evaluates their performances in handwritten character recognition. The algorithms include the well-known LVQ and those that globally optimize an objective function, as well as some newly derived variants. Experimental results of handwritten numeral recognition and Chinese character recognition show that the global optimization algorithms generally outperform LVQ. Particularly, the generalized LVQ of Sato and Yamada (1998) and a new algorithm MAXP2 yield best results.

The paper describes a relaxation based matching method for offline character recognition. This method employs elastic stroke models as standard character patterns. Pattern similarity between a standard and an input pattern is defined by fuzzy logic. The matching process is formalized as a maximization problem of the similarity and computed by the steepest descent technique. To implement this technique, we adopted automatic differentiation, which made it possible to calculate the partial derivatives of the target function automatically, only given the definition of that function. Results of computer experiments targeting 46 hiragana characters from the ETL8B database revealed a maximum recognition rate of 98.8% for 20 input sets when combining stroke springs with relative location springs.

This paper proposes a model-dependent pattern normalization method that employs an elastic membrane model to absorb pattern deformation in handwritten character patterns. This method prepares an elastic membrane from an input pattern and induces model-dependent pattern normalization by placing the membrane on each potential field representing a standard pattern. We composed a recognition system from non-linear normalization, coarse classification, model-dependent normalization and fine classification, then show that the system with the model-dependent normalization is more effective than that without it for both the ETL off-line handwritten character patterns database and the HANDS on-line patterns database.

Describes a handwritten character recognition system that integrates offline recognition requiring a bitmap image and online recognition involving an input pattern as a sequence of x-y coordinates. Offline recognition performs well for painted or overwritten patterns (for which online recognition would not be suited), whereas online recognition is suitable for very deformed patterns (for which offline recognition is not suited). Because each method has different recognition capabilities, the methods complement each other when integrated together. We have implemented a hybrid handwritten character recognition system in which the recognition results of the offline and online recognizer are integrated to create an improved product. After testing several integration methods for a handwritten character database, we found that the best method increased the recognition rate from 73.8% (offline) and 84.8% (online) to 87.6% (integrated).

The paper presents an offline handwritten character recognition method based on a dynamic system. The method employs rubber string models for standard character patterns (called an "elastic stroke model") and matches them with the potential field made from an input pattern. The matching process is realized by a dynamic relaxation with DP algorithm which minimizes the dynamic energy of rubber and potential field. To study the effect of the local minimum problem inherent in the relaxation technique, we used an improved DP method implemented on a parallel machine to compute the global minimum. Some experiments are performed on the 10 sets of 46 categories from the Japanese set of Hiragana characters that come from the ETL8b2c database. As a result, it is shown that the local minimum has little influence on recognition rates in this method, and the elastic stroke model could absorb deformation of character patterns.

The paper presents a database of on-line handwritten character patterns sampled in a sequence of sentences without any instructions. The sentences according to which character patterns are collected have been picked up from newspaper to include 1227 frequently appearing character categories with the result that they are composed of about 10000 characters and include 1537 JIS 1st level character categories. The rest of the JIS 1st level 1808 categories have been added at the end of the above text and written one by one. The total text has been commonly employed for collecting script patterns from a number of people. Patterns offered were inspected and omissions and wrong patterns were rewritten. The authors collected data from 80 people and made the 12000/spl times/80 patterns available from February 1996. More patterns are being collected. The paper describes the characteristics of this database as well as several tools to collect patterns.

This paper describes a new online handwritten character recognition system which is composed of coarse classification, linear-time elastic matching, structured character pattern representation and context post-processing. It has marked 90 to 95% correct recognition rates without learning to a large database of on-line handwritten Japanese text. The recognition time is about 0.3 sec./input character on an i486 DX2/66 MHz processor. The system is not only robust to pattern distortions but also highly customizable for personal use. Upon the request of learning an input pattern, it investigates which subpattern (radical) or the pattern as a whole is non-standard, registers the (sub)pattern and extends the effect of the registration to all the character categories whose shapes include it.

A simple and fast segmentation of text and line drawings which exploits the characteristics of online handwritten patterns is presented. Many attempts have been made on text/figure segmentation in offline research, but none in online. In online environment, mode selection is easily employed to input characters and drawings commongly. But it disturbs the user's thought. The authors propose online but delayed processing of handwritten patterns after they are written. Because the input patterns consists of characters and line drawings, there is a need to discriminate input patterns into characters and line drawings. An analysis of 17 handwritten patterns shows that the logarithmic distribution of character stroke length and that of the length of the longer side of rectangles circumscribing character strokes have similarity to the normal distribution. This finding, the contact frequency of character strokes and drawing strokes, and stroke input order have been used to perform the segmentation. In the testing of 18 patterns, 88% of all strokes were correctly judged.>

Basic design concepts are presented for a creative writing environment with pen and display integrated tablet to promote human thinking rather than to bother a user with the chores of input. The pen is suitable for creative work since one can express almost everything and is not bothered by the method of use. Experimental pen-based systems and products have not exploited the automated nature of handwriting. They try to recognized handwriting immediately after each pattern is written with the result of frequent misrecognition, and thus they interrupt the user's thinking. A structure of representation to handle the ambiguity of handwritten patterns is presented, as well as a lazy recognition scheme which delays the display of recognition until needed. Also presented is a prototype for text preparation, automatic segmentation of text and diagrams and mathematical formula recognition. They are all consolidated with lazy recognition.>

The authors present a handwritten character recognition method along with experiments that were conducted on a system developed to test this method's feasibility. This method can be applied to both online and offline input. Under this method the matching of characters is based on a dynamic model and performed using a relaxation method. A potential field is created to represent the dictionary character. Into this are dropped feature points taken from the input character and matching takes place. Matching thus occurs with the input character distorting to match the dictionary character. Springs are used to control this distortion. Details of variations to this method are given, and the results of comparative experiments conducted on these variations are presented.>

A programming philosophy for programming in a mother tongue, particularly in Japanese, implementation of a computer system where Japanese can be used for any purpose, the practice of programming in Japanese, and its effect are discussed. The key to the system of using Japanese is a language compiler for systems implementation. Software was developed for this system and a notable difference was observed.>