Search

Electroencephalography (EEG) captures the electrical activities of human brain. It is an easy and cost effective tool to characterize motor imager (MI) task used in brain computer interface (BCI) implementation. The MI task is represented by short time trial of multichannel EEG. In this paper, the raw EEG trial is regenerated using narrowband signals obtained from individual channel. Each channel of EEG trial is decomposed into a set of subband signals using multivariate discrete wavelet transform. The selected subbands are organized in two different ways namely vertical arrangement of subbands (VaS) and horizontal arrangement of subbands (HaS) to regenerate the trials. The features are extracted from each of the arrangements using common spatial pattern (CSP). An optimum number of features are used to classify the motor imagery tasks represented by EEG trials. The effectiveness of two classifiers- linear discriminant analysis (LDA) and support vector machine (SVM) are studied. The performances of the proposed methods are evaluated using publicly available benchmark datasets. The experimental results show that it performs better than the recently developed algorithms.

Handwriting analysis is playing an important role in user authentication or online writer identification for more than a decade. It has a significant role in different applications such as e-security, signature biometrics, e-health, gesture analysis, diagnosis system of Parkinson’s disease, Attention-deficit/hyperactivity disorders, analysis of vulnerable people (stressed, elderly, or drugged), prediction of gender, handedness and so on. Classical authentication systems are image-based, text-dependent, and password or fingerprint-based where the former one has the risk of information leakage. Alternatively, image processing and pattern-analysis-based systems are vulnerable to camera attributes, camera frames, light effect, and the quality of the image or pattern. Thus, in this paper, we concentrate on real-time and context-free handwriting data analysis for robust user authentication systems using digital pen-tablet sensor data. Most of the state-of-the-art authentication models show suboptimal performance for improper features. This research proposed a robust and efficient user identification system using an optimal feature selection technique based on the features from the sensor’s signal of pen and tablet devices. The proposed system includes more genuine and accurate numerical data which are used for features extraction model based on both the kinematic and statistical features of individual handwritings. Sensor data of digital pen-tablet devices generate high dimensional feature vectors for user identification. However, all the features do not play equal contribution to identify a user. Hence, to find out the optimal features, we utilized a hybrid feature selection model. Extracted features are then fed to the popular machine learning (ML) algorithms to generate a nonlinear classifier through training and testing phases. The experimental result analysis shows that the proposed model achieves more accurate and satisfactory results which ensure the practicality of our system for user identification with low computational cost.

Sign language is designed to assist the deaf and hard of hearing community to convey messages and connect with society. Sign language recognition has been an important domain of research for a long time. Previously, sensor-based approaches have obtained higher accuracy than vision-based approaches. Due to the cost-effectiveness of vision-based approaches, researchers have been conducted here also despite the accuracy drop. The purpose of this research is to recognize American sign characters using hand images obtained from a web camera. In this work, the media-pipe hands algorithm was used for estimating hand joints from RGB images of hands obtained from a web camera and two types of features were generated from the estimated coordinates of the joints obtained for classification: one is the distances between the joint points and the other one is the angles between vectors and 3D axes. The classifiers utilized to classify the characters were support vector machine (SVM) and light gradient boosting machine (GBM). Three character datasets were used for recognition: the ASL Alphabet dataset, the Massey dataset, and the finger spelling A dataset. The results obtained were 99.39% for the Massey dataset, 87.60% for the ASL Alphabet dataset, and 98.45% for Finger Spelling A dataset. The proposed design for automatic American sign language recognition is cost-effective, computationally inexpensive, does not require any special sensors or devices, and has outperformed previous studies.

The novel COVID-19 coronavirus has become an important threatening issue now for billions of humans. But lacking of available testing kits lead us to develop automatic detection system to prevent the spread of COVID-19. Since, most COVID-19 patients suffer the lung infection, chest X-ray can be used to be an effective imaging technique to detect COVID-19. Because of the limited number of COVID-19 chest X-ray images, it is efficient to use deep transfer learning that can provide a promising solution by transferring knowledge from generic object recognition task to domain-specific task. In this work, we aim to develop a deep transfer learning based detection of COVID-19 from chest X-ray images using imageNet pre-trained VGG-16 deep CNN model. We develop six binary classifiers for different chest X-ray image data and then integrate all the classifiers to produce a final classifier to predict COVID-19 from non COVID-19 cases (normal and other disease). To evaluate all the models, we have collected COVID-19 chest X-ray images from two open source GitHub repository and others chest X-ray images from RSNA Pneumonia Detection Challenge dataset. The experimental result shows the capability of our final classifier in the detection of COVID-19 by evaluating it with an independent testset. It also shows that our classifier achieved a high accuracy of 93% (with sensitivity of 87%, specificity of 94%, precision of 100% and F1-Score of 93%) in the detection of COVID-19 X-ray images from normal and other disease.

When the movement or musculoskeletal system (i.e., muscles, ligaments, nerves, etc.) of the human body is disrupted or affected through injury, it is called musculoskeletal disorders or MSDs. As the diagnosis of musculoskeletal conditions is complicated, detecting abnormalities from radiographs can be very significant in medical image analysis. Various computational methods have been introduced for this task which provided good classification accuracy. In this paper, we utilized the transfer learning approach to detect abnormalities because of its ability to share knowledge from similar tasks. We used an updated version of a pre-trained model (DenseNet169) to detect abnormalities for five different organs of the upper extremity. We applied data augmentation, resizing, and cropping for data preprocessing. We also tuned a couple of hyperparameters to improve our model’s performance. For evaluation, we calculated some well-known metrics to verify our model’s performance. We also compared our model’s performance with previous classifiers and found promising results. For one of the study types (finger), our classifier’s performance is improved by 67.05%. This proved that our model has the potential to be a useful tool for abnormality detection from radiographic images.

In developed countries, Alzheimer’s disease (AD) is one of the major causes of death. Until now, clinically there is not have any diagnostic method available but from a research point of view, this disease detection accuracy is produced by computational algorithms. There are many researchers who are working to find about Alzheimer’s disease property, its stages, and classification ways. This research plays a vital role in clinical tests for medical researchers and in the overall medical sector. One of the major problems found by the researchers in the field is the large data dimension. In the study, we proposed an efficient dimensionality reduction method to improve Alzheimer’s dis-ease (AD) detection accuracy. To implement the method first we cleaned the dataset to remove the null value and removing other unacceptable data by some preprocessing tasks. On the preprocessed data first, we have split into training and test dataset then we employed a dimension reduction method and there-fore applied a machine learning algorithm on the reduced dataset to produce accuracy for detecting Alzheimer’s disease. To overserve and calculate the accuracy we computed confusion matrix, precision, recall, f1-score value and finally accuracy of the method as well. Reducing the dimension of data here we applied consequently Principle component analysis (PCA), Random Projection (RP) and Feature Agglomeration (FA). On the reduced features, we have applied the Random Forest (RF) and Convolution neural network (CNN) machine-learning algorithm based on the dimensionality reduction method. To evaluate our proposed methodology here we have used Alzheimer’s disease neuroimaging initiative (ADNI) dataset. We have experimented for (i) Random forest with principal component analysis (RFPCA), (ii) Convolution neural network (CNN) with (PCA) (CNNPCA), (iii) Random forest with Ran-dom projection(RFRR), and (iv) Random forest with Feature agglomeration(RFFA) to differentiate patients with the AD from healthy patients. Our model namely Random forest with Random projection(RFRP) has produced 93% accuracy. We believe that our work will be recognized as a groundbreaking discovery in this domain.

In recent years, hand gesture recognition (HGR) systems have been extensively developed with the technologies of human-computer interaction (HCI), which enables regular interaction with machines. However, the significance of progress in HGR continues to advance, although the problem of hand segmentation and recognition is challenging due to the unfavorable environment, background illumination, hand size, and shape. To overcome this, we propose an optimal segmentation method for identifying hand gestures from input images, improving recognition performance. For segmenting hand gestures, we compared the segmentation methods of YCbCr, SkinMask, and HSV (hue, saturation, and value). The CR component is extracted from YCbCr, then binarization, erosion, and hole filling are performed. Color segmentation is applied to the SkinMask process that detects pixels that match the color of the hand. In the HSV process, threshold masking determines the dominant features. The Softmax classification is used to classify hand gestures where features are extracted through convolutional neural network (CNN). The proposed segmentation methods are applied to a benchmark dataset and the result shows an improvement in recognition accuracy over state-of-the-art systems.

Identifying a user or a person through handwriting-data is a popular technique. Many researches have been done in this area most of which are image or pattern-based analysis. The accuracy level depends on the quality of the image or pattern. In this paper, we proposed user authentication system using an individual's pen tablet handwriting data. The proposed system is concerned with the numerical value of a person handwriting data getting from the digital pen and tablet device. Hence, user authentication through pen tablet data ensures more accuracy by working with user's real time handwritten data. In the proposed system, 24 persons writing samples(1262 .csv files and 23 class)are used for extracting features to identify a user based on their handwriting attributes. Six completely separated features are extracted after data analysis and pre-processing. The extracted features are mainly concerned with the vital attributes of a user's handwriting. The extracted features are used for classification. With this concern, we utilized different classification algorithms such as Support Vector Machine (SVM), Logistic Regression (LR), Linear Discriminant Analysis (LDA) and Random Forest (RF) classifier. From the implementation, different algorithms show different accuracy level. The testing accuracy rate of SVM, LR, LDA and RF is 87%, 85%, 76% and 77% respectively. The experimental analysis shows that we got more robust and satisfactory results which ensure the practicality of our system.

This article concerns smoke detection in the early stages of a fire. Using the computer-aided system, the efficient and early detection of smoke may stop a massive fire incident. Without considering the multiple moving objects on background and smoke particles analysis (i.e., pattern recognition), smoke detection models show suboptimal performance. To address this, this paper proposes a hybrid smoke segmentation and an efficient symmetrical simulation model of dynamic smoke to extract a smoke growth feature based on temporal frames from a video. In this model, smoke is segmented from the multi-moving object on the complex background using the Gaussian&rsquo, s Mixture Model (GMM) and HSV (hue-saturation-value) color segmentation to encounter the candidate smoke and non-smoke regions in the preprocessing stage. The preprocessed temporal frames with moving smoke are analyzed by the dynamic smoke growth analysis and spatial-temporal frame energy feature extraction model. In dynamic smoke growth analysis, the temporal frames are segmented in blocks and the smoke growth representations are formulated from corresponding blocks. Finally, the classifier was trained using the extracted features to classify and detect smoke using a Radial Basis Function (RBF) non-linear Gaussian kernel-based binary Support Vector Machine (SVM). For validating the proposed smoke detection model, multi-conditional video clips are used. The experimental results suggest that the proposed model outperforms state-of-the-art algorithms.

As it is the seventh most-spoken language and fifth most-spoken native language in the world, the domain of Bengali handwritten character recognition has fascinated researchers for decades. Although other popular languages i.e., English, Chinese, Hindi, Spanish, etc. have received many contributions in the area of handwritten character recognition, Bengali has not received many noteworthy contributions in this domain because of the complex curvatures and similar writing fashions of Bengali characters. Previously, studies were conducted by using different approaches based on traditional learning, and deep learning. In this research, we proposed a low-cost novel convolutional neural network architecture for the recognition of Bengali characters with only 2.24 to 2.43 million parameters based on the number of output classes. We considered 8 different formations of CMATERdb datasets based on previous studies for the training phase. With experimental analysis, we showed that our proposed system outperformed previous works by a noteworthy margin for all 8 datasets. Moreover, we tested our trained models on other available Bengali characters datasets such as Ekush, BanglaLekha, and NumtaDB datasets. Our proposed architecture achieved 96–99% overall accuracies for these datasets as well. We believe our contributions will be beneficial for developing an automated high-performance recognition tool for Bengali handwritten characters.

Signature verification is one of the most popular subjects in pattern recognition. Many kinds of verification methods for online handwritten signatures have basically used the individual features of signatures. However, how to reflect inter-stroke information extracted from handwritten multi-stroke signatures in the verification has not been well considered. This paper suggests a new verification method that uses inter-stroke information based on shape contexts. The shape context describes how points are distributed around a given point on the shape. We compare our Shape Context method with a basic method. The basic method performs signature verification using intra-stroke information such as pen position, pen pressure, pen inclination, pen altitude and elapsed time by DP matching. The Shape Context method treats inter-stroke information such as point distribution. In addition, we try to improve the accuracy of the method by incorporating a weighted evaluation of the average pressure value of each stroke. Comparing with the basic method, the Shape Context method reduces false rejection rate from 4.14% to 1.54%, and false acceptance rate from 4.06% to 2.01%. The experimental results show the effectiveness of inter-stroke information on signature verification.

This paper deals with the problem of fault detection and diagnosis in sensors considering erratic, drift, hard-over, spike, and stuck faults. The data set containing samples of the abovementioned fault signals was acquired as follows: normal data signals were obtained from a temperature-to-voltage converter by using an Arduino Uno microcontroller board and MATLAB. Then, faults were simulated in normal data to get 100 samples of each fault, in which one sample is composed of 1000 data elements. A support vector machine (SVM) was used for data classification in a one-versus-rest manner. The statistical time-domain features, extracted from a sample, were used as a single observation for training and testing SVM. The number of features varied from 5 to 10 to examine the effect on accuracy of SVM. Three different kernel functions used to train SVM include linear, polynomial, and radial-basis function kernels. The fault occurrence event in fault samples was chosen randomly in some cases to replicate a practical scenario in industrial systems. The results show that an increase in the number of features from 5 to 10 hardly increases the total accuracy of the classifier. However, using ten features gives the highest accuracy for fault classification in an SVM. An increase in the number of training samples from 40 to 60 caused an overfitting problem. The $k$ -fold cross-validation technique was adopted to overcome this issue. The increase in number of data elements per sample to 2500 increases the efficiency of the classifier. However, an increase in the number of training samples to 400 reduces the capability of SVM to classify stuck fault. The receiver operating characteristics curve comparison shows the efficiency of SVM over a neural network.

Sleep disorders have a significant effect on psychological depression and many other human diseases. Nowadays, technology, as well as innovation, has become an essential and analytical part of the world. Detection of sleep disorders by brain waves has become a dynamic study. From this point of view, this paper proposed a model for detecting sleep disorders using time-frequency analysis based feature extraction model based on EEG (Electroencephalogram) signal. In this proposed method, Empirical mode decomposition (EMD) and wavelet packet transform (WPT) time-frequency analysis techniques are used for extracting effective features, because those techniques are very effective for analyzing the non-stationary signal like EEG. In this research, the EMD and WPT are used to decompose the input signal. In EMD decomposition, up to 9th IMFs (Intrinsic Mode Functions) are decomposed. In WPT decomposition, the EEG signal is decomposed up to third level wavelet coefficient. After the decomposition process, different statistical features are extracted, i.e., Shannon entropy, energy, standard deviation, skewness, and kurtosis. However, identifying the optimal sub-band of time-frequency analysis is very challenging. Thus, the genetic algorithm (GA) is used to select the effective subset of the feature. In the detection process, SVM classifier is used and sleep disorders are classified based on trained knowledge. As a result, the performance of the proposed method is evaluated for various statistical features and to find the optimal features for detecting sleep disorder. According to the experimental results, the proposed model shows improved performance by 4.88% improved classification accuracy.

Motor imagery (MI) from human brain signals can diagnose or aid specific physical activities for rehabilitation, recreation, device control, and technology assistance. It is a dynamic state in learning and practicing movement tracking when a person mentally imitates physical activity. Recently, it has been determined that a brain&ndash, computer interface (BCI) can support this kind of neurological rehabilitation or mental practice of action. In this context, MI data have been captured via non-invasive electroencephalogram (EEGs), and EEG-based BCIs are expected to become clinically and recreationally ground-breaking technology. However, determining a set of efficient and relevant features for the classification step was a challenge. In this paper, we specifically focus on feature extraction, feature selection, and classification strategies based on MI-EEG data. In an MI-based BCI domain, covariance metrics can play important roles in extracting discriminatory features from EEG datasets. To explore efficient and discriminatory features for the enhancement of MI classification, we introduced a median absolute deviation (MAD) strategy that calculates the average sample covariance matrices (SCMs) to select optimal accurate reference metrics in a tangent space mapping (TSM)-based MI-EEG. Furthermore, all data from SCM were projected using TSM according to the reference matrix that represents the featured vector. To increase performance, we reduced the dimensions and selected an optimum number of features using principal component analysis (PCA) along with an analysis of variance (ANOVA) that could classify MI tasks. Then, the selected features were used to develop linear discriminant analysis (LDA) training for classification. The benchmark datasets were considered for the evaluation and the results show that it provides better accuracy than more sophisticated methods.

Sleep disorders has a vital effect on mental depression and many other diseases of human body. Diagnosing the sleep disorder in an early curable stage may help to provide better treatment and save the life. The EEG (Electroencephalogram) signal is one of the most uses bio signal for capturing brain activities to detect and diagnosis the sleep disorders. Empirical mode decomposition (EMD) is an efficient time-frequency data analysis technique for diagnosing disease by analyzing EEG signal. However, it is a challenging issue to select the optimal intrinsic mode functions (IMFs) of Empirical mode decomposition (EMD) for extracting discriminant properties of EEG signals to diagnosis the sleep disorder. From this point of view, this paper presents a model to select optimal IMF of EMD for diagnosing the sleep disorder using EEG brain signal. In this proposed model, EMD is applied to decompose and analyze EEG signal for extracting biomarker/feature of sleep disorders. During the EMD decomposition process, different levels of IMF are extracted and features, i.e., Shannon Entropy, Spectral Entropy, Standard deviation, Skewness and Kurtosis are calculated from those IMFs for detecting the sleep disorders. In identification process, the multiclass support vector machine (MC-SVM) classification algorithm is used and sleep disorders are classified based on trained knowledge. Finally, the performance of proposed model is evaluated for different IMFs of EMD and find the optimal IMF for sleep disorder diagnosis. For evaluating the proposed model, a benchmark dataset including 4 types of data such as Apnea, REM, PLM and healthy subjects are used in experiment. According to the experimental result, the proposed model achieves the optimal classification performance for IMF 8, i.e., 93.24% average classification accuracy.

Various verification methods for on-line handwritten signatures have been proposed; however, reflecting inter-stroke information in the verification has not been considered. Hence, this paper suggests a verification method that uses Shape Context. In this paper, the verification method is divided into two methods. One is the basic method and the other is the Shape Context method. The basic method treats x-y coordinates, pen-pressure, pen-inclination, altitude and writing time and the Shape Context method treats inter-stroke information as point distribution. Also, the Shape Context method does the verification by incorporating a weighted evaluation of the average pressure value at each stroke. In the comparison experiment, FRR decreases from 4.14% to 1.54%, and FAR decreases from 4.06% to 2.01%. The experimental results show the effectiveness of the inter-stroke information method.

In this paper a new interactive map search system is presented using shape context and bipartite graph matching. Shape context is used for measuring shape similarity and the recovering of point correspondences. After the above information is generated from the shape context bipartite graph matching is used to obtain the optimal correspondence between two shapes. Hierarchical description is also used to increase the recognition rate. Shape context is a method to treat shapes as a set of points and generate the histogram of the distribution of points. Wavelet analysis is used in hierarchical description. In order to shorten the calculation time, piecewise linear approximation is implemented as the feature extraction method. The systemlists the sixmost similar shapes to hand-written input shapes from the reference shapes, i.e., Japan’s 47 prefectures. Comparison results of linear matching, Dynamic Programming (DP) matching and shape context with bipartite graph matching indicate that the 1st place recognition rates are 82%, 84.52% and 92.45%, respectively. The evaluation result of hierarchical description shows that hierarchical approximation can improve the recognition rate from 92.45 to 94.97% using the deepest-4 depth. These results show that the proposed method is effective on fulfilling the interactive map search system.

This article presents a new method for generating random handwriting fonts. Random fonts are rendered differently every time they are output onto the screen. When generating handwriting fonts, it is important to render the individual’s handwriting style as this can facilitate and improve communication. The purpose of this research is to develop a system that simulates an individual’s handwriting style using randomly generated fonts. This method utilizes three main steps: dynamic programming (DP) for matching the original data points of input strokes, a minimum spanning tree method, and random font generation through interpolation. DP matching improves the accuracy of the minimum spanning tree. The minimum spanning tree connects similar shapes in order to produce a new font shape had original features truly. The interpolation method allows a generated font shape to be similar to real handwriting characters but have a new shape. As a result, our system generates random fonts that retain an individual handwriting features, maintaining the wealth of font variation.

As a means to perform on-line recognition of cursive Korean characters, called hanguls, we describe a structural analysis type algorithm that searches globally for key points of segmentation on a character unit level and can cope with large variations in stroke shape and position. This "segmentation points search" is systematically performed by a two-level dynamic programming (DP) matching algorithm in conjunction with syntax control of hangul composition characteristics. Fine discrimination for phonemes and characters is effectively realized using mutual information among strokes. Experiments demonstrate computational feasibility and that the proposed approach provides high recognition and segmentation ability.

This paper describes an optimal stroke-correspondence search method that makes possible stroke-order-free on-line character recognition. During the stroke-correspondence search process, conventional individual stroke-information regarding the shape and position of each stroke, and interstroke-information regarding the mutual relationships among strokes are both employed. The optimal path search for stroke-correspondence, being based on an optimal criterion including both intra- and interstroke-information, is systematically carried out. The reasonable level stroke-correspondence search is achieved partially by using the information regarding the actually occurring stroke-order, which does not hinder the framework of stroke-order-free recognition due to its use of statistically stable information. A large improvement in both computational time and recognition-accuracy were achieved in the current experiments.

In this paper, the features that have personal characteristic using pen inclination and pressure information are discussed. Forging a pen inclination and pressure information is difficult because it is not visible. Four features using invisible information are proposed and their characteristics are discussed. Proposed features calculated by physical vector analysis are verified by SVC2004 database using DP matching algorithm. As a result, the new feature named Down improves the recognition rate and reliability. Average of correct verification rate is 94.57 % and variance is 0.667. DOI: http://dx.doi.org/10.11591/ijece.v2i4.474

Signature evaluation is a method to determine individual signature whether that is good or bad. A good signature is more complex and difficult to impersonate, besides a bad signature are easy to recognition and easy to impersonate. If a signature have a lot of angle, that is means the signature is more complex, so we calculate the internal angle quantity to determine that is a good or bad signature. There are too many angle in a signature, contain big angle or small angle. In this paper, we using geometry to decide what kind of angle that we want to determine.

One essential task in extracting information from biomedical literature is the bio Named Entity Recognition (NER) process, which basically defines the boundaries between typical words and biomedical terminology in particular text data, and assigns them based on domain knowledge. This paper presents a semi supervised integration of completely different classifiers to cover knowledge from unlabeled data to recognize bio named entities in text. We modified the original co-training, a semi supervised learning algorithm, with a scalable feature processing schema, which extracts the bio NER feature from a number of unlabeled data and converts different types of feature sets. Our base result shows that the classifiers of co-training achieve significant learning from unlabeled data.

This paper presents a map search system using Shape context and Bipartite graph matching. Shape context is using for measuring shape similarity and the recovering of point correspondences. After generated from the shape context, Bipartite Graph Matching leads the optimal correspondence between two shapes. For raising the recognition rate, Hierarchical description is used. Shape context is the method that treats shapes as a set of points and makes the histogram of the distribution of points. In Hierarchical description, Wavelet analysis is used. To reduce the amount of calculation, Piecewise linear approximation is implemented as the feature extraction method. The system shows six similar shapes to hand-written input shapes from reference shapes that are Japan's 47 prefectures. Comparison result of Linear matching, DP matching and Shape context with Bipartite graph matching describes that the 1st place recognition rates of each algorithms are 82.00%, 84.52% and 92.45%, which describes the robustness of Shape context. Experiment of several level of hierarchical description describes that hierarchical approximation can raise the recognition rate from 92.45 to 94.97% using the deepest-4 depth, which means that the hierarchical approximation is effective to remove the noise from the handwritten inputs.

A lot of image data has been digitized and preserved in computers. In order to search for an image, an efficient and accurate retrieval method is needed. This paper is concerned with shape retrieval which is one of the searching methods for finding similar images in a database. Shape, the outer form of a picture, is considered the most promising feature to identify entities in an image. The problem of shape retrieval takes a lot of time because an exhaustive search is mainly used in literatures. This paper suggests the use of a clustering method known as dendrogram for shape retrieval. In addition, we proposed the automatic decision of a threshold to determine a number of clusters in dendrogram. Through the experimental result, the proposed method proved fast retrieval preserving almost the same level accuracy.

This paper considers the unsupervised classification of handwritten character strokes in regards to speed, since handwritten strokes prove challenging with their high and variable dimensions for classification problems. Our approach employs a robust feature detection method for brief classification. The dimension is reduced by selecting feature points among all the points within strokes, and thus the need to compare stroke signals between two different dimensions is eliminated. Although there are some remaining problems with misclassification, we safely classify strokes according to handwriting styles through a refinement procedure. This paper illustrates that the equalization problem, the severe difference in small parts between two strokes, can be ignored by summing all of the differences via our method.

In this paper, we refer a phenomenon of relatively high error rates against random forgeries in online signature verification. Though human can easily distinguish random forgeries from genuine signatures, the most published research results during last decades have reported high error rates against them, compared to the error rates against skilled forgeries. This paper suggests taking DPW (dynamic positional warping) approach for the problem, which supplies high accuracy and robustness against affine transform of online shapes. In addition, the effects of distance normalization and slope constraints are tested to DPW and conventional DTW (dynamic time warping). Our approach was compared to conventional DTW method with consequential xy difference information with SVC2004 task1 database: 0.17% and 1.47% of equal error rates are reported against random and skilled forgeries, respectively. Comparing to the best result of the conventional approach, the proposed approach reduced the error rates into about 27% in a single procedure without much additional time costs.

In the stroke-based signature verification system, it is important to find correct points of a testing signature to be spilt according to its template signature. In this paper, we propose a modified dynamic time warping algorithm (DTW) for the problem. Proposed method has two differences to the conventional approaches. First, the stroke information of both template and testing signature is considered precious to prevent wrong splitting. Second, the number of stroke difference in genuine signature is not limited because the proposed method does not require much additional computational time. Through the experiment, proposed method is verified with 1359 signatures written by 17 Japanese writers. The test data were all the semi-skilled forgeries, and the results of proposed methods showed 3.85% in the equal error rate (the error rate when the false acceptance rate and false rejection rate is the same; EER) while the EER of the compared method was 7.8%.

This paper reports on a compression method for a reference character dictionary used for on-line Chinese handwriting character recognition, which is based on pattern matching between input and reference patterns. First, one reference pattern for each category is generated from the training data. Second, the reference dictionary is optimized by using the Generalized Learning Vector Quantization (GLVQ) algorithm. Third, the reference dictionaries are compressed to approximately one tenth of the original data using Vector Quantization with the K-means clustering algorithm. Then, the compressed reference dictionaries are again optimized by the GLVQ algorithm. Experimental results for Chinese character recognition show that the dictionary can be successfully compressed without decreasing the recognition rate, and the calculation time of distance between strokes can be reduced by a factor of approximately five.

In this paper a new algorithm for the on-line recognition of cursive Korean characters (hangul) is investigated for the problem of searching the segmentation points to divide the phonemes. The segmentation points searching is systematically performed by two-level DP matching algorithm, that is, one algorithm targets the phoneme unit level, and the other the whole character unit level. Based on the composition characteristics of hanguls, effective syntax is introduced as part of the character level process. Fine discrimination is effectively realized for both phonemes and characters including the individual information of each stroke and mutual information among strokes. Experiments demonstrate real time feasibility, high recognition and segmentation performance.