Search

We propose a method to detect human wants by using an electroencephalogram (EEG) test and specifying brain activity sensing positions. EEG signals can be analyzed by using various techniques. Recently, convolutional neural networks (CNNs) have been employed to analyze EEG signals, and these analyses have produced excellent results. Therefore, this paper employs CNN to extract EEG features. Also, support vector machines (SVMs) have shown good results for EEG pattern classification. This paper employs SVMs to classify the human cognition into “wants,” “not wants,” and “other feelings.” In EEG measurements, the electrical activity of the brain is recorded using electrodes placed on the scalp. The sensing positions are related to the frontal cortex and/or temporal cortex activities although the mechanism to create wants is not clear. To specify the sensing positions and detect human wants, we conducted experiments using real EEG data. We confirmed that the mean and standard deviation values of the detection accuracy rate were 99.4% and 0.58%, respectively, when the target sensing positions were related to the frontal and temporal cortex activities. These results prove that both the frontal and temporal cortex activities are relevant for creating wants in the human brain, and that CNN and SVM are effective for the detection of human wants.

With the progress of IoT, everything is going to be connected to the network. It will bring us a lot of benefits however some security risks will be occurred by connecting network. To avoid such problems, it is indispensable to strengthen security more than now. We focus on personal authentication as one of the security. As a security enhancement method, we proposed a method to carry out numeral identification and personal authentication using numerals written in the air with Leap motion sensor. In this paper, we also focus on proper handling of aerial input numerals to verify whether the numerals written in the air are helpful for authentication. We collect numerals 0 to 9 from five subjects, then apply three pre-processing to these data, learn and authenticate them by CNN (convolutional neural network) which is a method of machine learning. As a result of learning, an average authentication accuracy was 92.4%. This result suggests that numerals written in the air are possible to carry out personal authentication and it will be able to construct a better authentication system.

This paper proposes a method to classify words in Japanese Sign Language (JSL). This approach employs a combined gathered image generation technique and a neural network with convolutional and pooling layers (CNNs). The gathered image generation generates images based on mean images. Herein, the maximum difference value is between blocks of mean and JSL motions images. The gathered images comprise blocks that having the calculated maximum difference value. CNNs extract the features of the gathered images, while a support vector machine for multi-class classification, and a multilayer perceptron are employed to classify 20 JSL words. The experimental results had 94.1% for the mean recognition accuracy of the proposed method. These results suggest that the proposed method can obtain information to classify the sample words.

In this paper, we propose to measure the EMGs of the wrist and fingers using dry-type sensors worn near the wrist, and to separate the measured data into wrist and finger EMGs by using independent component analysis (ICA). Then we can confirm the EMGs of the wrist and fingers from the complex motion and realize individual identification in more complex motions. The final goal of this study is to identify individual motions from complex motions. In this paper, as a preliminary step, the ICA is used to isolate compound motions and the validity of the method is evaluated. We measured the EMGs for three days and four motions. The results of the combination of FastICA, Infomax and JADE, respectively, were evaluated by the correlation coefficient with the original signal. The most accurate combination was FastICA + Infomax with an accuracy of 70.5%

We use Leap Motion and a deep neural network to perform personal authentication and character recognition of all hiragana characters entered in the air. We use Leap Motion to detect the index finger and store its trajectory as time series data. The input data was pre-processed to unify the data length by linear interpolation. For identification, the accuracy of Long Short Term Memory (LSTM) was compared with Support Vector Machine (SVM). As a result, SVM and LSTM achieved 97.25% and 98.18% F-measure in character recognition, respectively. In personal authentication, SVM has an accuracy of 92.45%, False Acceptance Rate (FAR) was 0.73%, and False Rejection Rate (FRR) was 41.59%. On the other hand, LSTM had an accuracy of 96.13%, FAR of 1.73% and FRR of 14.55%. Overall, the LSTM performed better than the SVM.

Recent years, interfaces using biometric information are progressing. Electromyogram(EMG) has been used in a variety of situations. Many studies have measured EMG in the shoulders and arms, where there is a lot of muscle mass. In addition, wet type sensors have been often used. However, those are inconvenient to use in everyday life and high cost. In this research, we measure wrist EMG for convenience and cost. Currently, researches have been done on the wrist EMG motion identification and personal identification. These studies have conducted simple movements and a large number of electrodes for discrimination. Furthermore, authentication by password sequence with gestures has not been done. In this paper, we propose to realize motion identification and personal authentication with complex movements using a small number of electrodes. The measured data was preprocessed such as removing noise and smoothing. We compared the accuracies obtained using Support Vector Machine(SVM) and Long Short-term memory(LSTM) for motion identification and authentication. The accuracies obtained using SVM and LSTM were 60.4% and 62.4%, respectively. In this case, the number of data was small. It is therefore necessary for increasing the number of data to perform deep learning.

This paper proposes a method to evaluate a sound sleep using an image gathering technique and its analysis techniques. The proposed method consists of three phases; gathered images generation, gathered images analysis and sound sleep evaluation. The gathered images designed to gather sleep postures and their changes are generated at 1 second, 10 seconds, 1 minute, 10 minutes, 1 hour intervals and all times, respectively. In the gathered image analysis, the gathered images are analyzed by calculating difference values among the gathered images of 10-minute and all times. Then, the sound sleep conditions are evaluated by visual inspection and analysis results. In order to show the effectiveness of the proposed method, we conduct experiments using real movies and their images. In experimental results, we confirm that there were sound sleep conditions, bad sleep conditions and borderline cases by checking subjective evaluation using questionnaire and generated gathered images visually. Moreover, we confirm that the calculated difference values among the gathered images of 10-minute and all times are different between sound sleep and other cases. Furthermore, the analyzed results show that the proposed method was successful in the sleep conditions classifications on four of five subjects. These results suggest that the gathered images analysis method is effective for evaluating whether sleep condition is sound sleep or not. In particular, it is important to calculate the difference values among the gathered images of 10-minute and all times to evaluate sleeping conditions.

A method to classify Japanese sign language (JSL) words using a gathered image generation method and a convolutional neural network (CNN) is proposed. The JSL words consist of words that are often used in information queries. Gathered images are generated based on the difference between the first image, which indicates the start position of a JSL word and target images, which indicate the motion position for a JSL word. The CNN is used to extract features from the gathered images. The JSL words are classified using a support vector machine. To show the effectiveness of the proposed method, we conducted experiments and computer simulations. We confirmed that the mean recognition accuracy for 10, 20, and 42 JSL words was 99.2%, 94.3, and 86.2%, respectively.

This paper introduces an electroencephalogram (EEG) analysis method to detect preferences for particular sounds. Our study aims to create novel brain–computer interfaces (BMIs) to control human mental (NBMICM), which are used to detect human mental conditions i.e., preferences, thinking, and consciousness, choose stimuli to control these mental conditions, and evaluate these choices. It is important to detect the preferences on stimuli. If the stimuli related to the preference can be detected, the NBCIMC can provide stimuli to the user based on their emotions by detecting their favorite stimuli. The proposed method adopted EEG recording technique, extraction techniques of EEG features and detection methods of preferences. EEG recording employs a simple electroencephalograph, for which the measurement position is the left frontal lobe (Fp1) of the brain. We assume that the differences of the EEG activities on the patterns of preference are expressed in the association between the changes of the power spectra on each frequency band of the EEG. To calculate the association, we employ the gray theory model. The EEG feature is extracted by calculating the gray association degree, then, the preferences are detect using a support vector machine (SVM). Experiments are conducted to test the effectiveness of this method, which is validated by a mean accuracy rate >88% on the favorite sound detection. These results suggest that the detection of subject’s favorite sounds becomes easy when the EEG signals are analyzed while the gray associate degrees are used as the EEG feature and the SVM is used as the classifier.

This paper proposes a method to enter characters and/or character string in an augmented reality using a gesture motion. The proposed method detects the region of character string using the gesture motion. It consists of five phases; template generation, skin color detection, hand region detection, gesture motion extraction and designation of character string region. The template image consists of two fingers because a gesture is to take hold the tips of the first and second fingers. In the skin color detection, we extract the skin color on the basis of values in saturation by using threshold processing. The hand region is detected by calculating areas and detecting the area with the maximum value as a hand. The gesture motion is extracted using template matching. In order to show the effectiveness of the proposed method, we conduct experiments for character string specification.

In recent years, biological signals have attracted attention as tools for human interfaces. Researches on biological signals have been actively conducted. In this paper, we propose a method which distinguishes ten motions, such as “One” “Two” “Three” “Four” “Five” “Six” “Seven” “Eight” “Nine” and “Ten” by measured the electromyogram of the wrist. We measure data by installing 8 dry type sensors on the right wrist. We carry out frequency analysis using FFT and try to take 3 kinds of methods to remove noise. Finally, we use Support Vector Machine (SVM) for identification and classification. We conducted experiments with four subjects. In the experimental result, the accuracy of finger motions recognition was 65%. In the future, we will also add more methods to remove noise, and try to find other methods to improve the accuracy in the research.

Security cameras have increased in public facilities. The number of crimes has decreased by security cameras, but we will have too much data of cameras. In this paper, we have aim of security improvement. First, we search whether there are humans in images by OpenPose. We then obtain position of human's hands. Finally, we detect dangerous objects around the hands by image classification.

In recent years, with the rapid development of the information society, the importance of personal authentication has become higher and higher. This paper focuses on the use of a Kinect sensor to obtain walking characteristics for personal authentication. In terms of the proposal method, Kinect is used to obtain body's physical feature quantity, such as the angle of joint bending when a person walks, the displacement of coordinates. In terms of learning recognition, the support vector machine and the obtained feature amount are used for personal authentication. We measured 3 subjects data 5 times a day for 4 days, and obtained an average recognition accuracy of 77.4 % using cross-validation.

In this paper, we propose a nail image analysis method to evaluate accumulated stress using fuzzy reasoning. The proposed method consists of three stages: measurement, feature extraction, and stress evaluation. In the measurement, we take a nail image. In the feature detection, we extract the lunula of the nail to calculate its height. In the stress evaluation, we evaluate accumulated stress using fuzzy reasoning. The experimental results suggest that the proposed method can determine the presence or absence of accumulated stress.

This paper proposes a method for classifying Japanese sign language (JSL) using a combined gathered image generation technique and a convolutional neural network (CNN) approach. In the combined gathered image generation, the maximum difference from the previous and next images is calculated for each block, and the block information that had maximum difference was embedded into an image on all blocks. After information on all images has been gathered into single words, the CNNs are used to extract features for the classification of JSL words. A multi-class support vector machine (SVM) is then used to classify words related to greeting and requesting. The mean and the standard deviation of the recognition accuracy of the proposed method were experimentally shown to be 84.2% and 4%, respectively. These results suggest that it is possible to obtain information for classifying 10 JSL words using the proposed combined gathered image generation and CNN approach.

As information technology has advanced in recent years, services which include personal authentication systems such as ATM are increasing. Current main personal authentication systems include IC cards, passwords, and biometrics authentication such as fingerprint authentication. However, there are several problems in these systems. Therefore, better systems are needed. As such systems, we propose a method to write numerals in the air using the Leap motion and to carry out personal authentication from such airlial handwriting data. We try to identify numerals 0 to 9 as a previous stage. After applying some pre-processing to inputs, learning and identification are carried out using CNN which is a method of machine learning. As a result, avarage identification accuracy rate was 93.4%. From this result, it is suggested that input numerals in the air can be identified and there is a possibility to constract a new personal authentication system.

As an alternative to voice, sign language and artificial larynx can be used. However, there are disadvantages where they require a long-term training and are expensive. Therefore, researches on detection of utterance by electromyography (EMG) analysis around the lips have been conducted. On the one hand, it is necessary to construct a personal authentication system to identify speakers. The electrode used in this paper is 2 electrodes sensor, which is small in size and a dry type. Three sensors are attached in the orbicularis muscle, the zygomatic major muscle, and the depressor angle oris muscle which can acquire myoelectric information necessary for identification in Japanese vowel utterance. EMG signals are measured using P-EMG plus. In order to eliminate noises, signal cutting is carried out before and after the central point of the acquired raw data. Furthermore, EMG data are divided to increase the number of data while overlapping. These are named “DATA 1”. A Hamming window is then applied for them, and the amplitude values of the power spectra are calculated by fast Fourier transform. Automatic verification and elimination of noise parts by quartile method were carried out. In order to reconstruct signals after noise elimination, the inverse Fourier transform is carried out and then a inverse Hamming window is applied. These are named “DATA 2”. Learning identification is carried out using a convolutional neural network. A large difference was found in accuracy depending on the data set created separately by measurement date. Therefore, it was found that intra-individual variation by each subject was large. In the future, it is necessary to further improve the data and to reduce individual variation within each subject.

In this paper, we propose an object search method which is adapted to transformation of an object to be searched to detect lost property. Object search is divided into two types; global and local searches. We used a template matching using Genetic Algorithm (GA) in the global search. Moreover we use a random search in the local search. According to experimental results, this system can detect rough position of the object to be searched. The search accuracy obtained using the present method is 83.6%, and that of a comparative experiment using only GA is 42.1%. We have verified that our proposed method is effective for lost property detection. In the future, we need to increase search accuracy to search objects more stably. In particular, we need to improve local search.

This paper introduces an electroencephalogram (EEG) analysis method to detect human preference. The proposed method consists of three phases; EEG recording, EEG feature extraction and preference detection. In EEG recording, we employ the simple electroencephalograph. The measurement position to record the EEG is left frontal lobe (FP1). The gray association degree is used to extract the EEG feature. The support vector machine is used to detect human preference on sounds listened to. In order to show the effectiveness of the proposed method, we conduct the experiments. In the experimental results, the mean of the accuracy rate of the favorite sound detection was higher than 88%.

This paper proposes a method to check whether human understands contents of learning using electroencephalogram (EEG). The proposed method consists of three stages; EEG measurement, EEG feature extraction and EEG pattern classification. The EEG feature extraction method (stage) consists of three phases; frequency analysis, band division based on rhythm of brain activities and principal component extraction. The frequency analysis converts an original waveform of EEG into frequency components (bands). In the band division, frequency components were divided into five rhythms (θ , Low-α , High-α , Low-β and High-β waves). In the principal component extraction, we calculate the first and the second principal components using the principal component analysis (PCA). In the EEG pattern classification, support vector machine (SVM) is used to classify EEG patterns based on PCA results. In order to show the effectiveness of the proposed method, we conduct experiments using real EEG data. The experimental results suggest that the most important sensing position to record EEG is F7 when checking whether he/she understands contents of learning using EEG.

Recognition of eye motions has attracted more and more attention of researchers all over the world in recent years. Compared with other body movements, eye motion is responsive and needs a low consumption of physical strength. In particular, for patients with severe physical disabilities, eye motion is the last spontaneous motion for them to make a respond. In order to provide an efficient means of communication for patients such as ALS (amyotrophic lateral sclerosis) who cannot move even their muscles except eye, in this paper we proposed a system that uses EOG signals and Learning Vector Quantization algorithm to recognize eye motions. According to recognition results, we use API (application programming interface) to control cursor movements. This system would be used as a means of communication to help ALS patients.

Recent years, biometrics authentication is receiving attention by development of information society. In this paper, we propose a personal authentication system, which uses behavior characteristics among biometrics. We focus on aerial handwritten signature, because it is difficult to forge it, and there is no risk of loss. In this paper, we measure signatures using Leap Motion Controller. It can measure three dimensional space coordinates with high accuracy. We divide signature data into three axial directions of coordinates XYZ in order to use them as one-dimensional data. We carry out preprocessing to signature data and normalize them. Next, we use deep learning based on a convolutional neural network for feature extraction and identification. In this experiment, we prepare data obtained from six subjects. We obtain genuine data of one subject. The remaining five subjects are used to create forgery data. We classify signature data into two classes. We conduct deep learning in which CNN carries out 10,000 cycles learning in one trial. We carry out this trial 5 times and evaluate mean accuracy by cross validation for two types of genuine data. The average discrimination accuracy of this experiment are 97.0 % and 95.9%. In addition, the false rejection rates are 9.6% and 19.2%. The false acceptance rate are 0.8% and 0.1%. Recent years, biometrics authentication is receiving attention by development of information society. In this paper, we propose a personal authentication system, which uses behavior characteristics among biometrics. We focus on aerial handwritten signature, because it is difficult to forge it, and there is no risk of loss. In this paper, we measure signatures using Leap Motion Controller. It can measure three dimensional space coordinates with high accuracy. We divide signature data into three axial directions of coordinates XYZ in order to use them as one-dimensional data. We carry out preprocessing to signature data and normalize them. Next, we use deep learning based on a convolutional neural network for feature extraction and identification. In this experiment, we prepare data obtained from six subjects. We obtain genuine data of one subject. The remaining five subjects are used to create forgery data. We classify signature data into two classes. We conduct deep learning in which CNN carries out 10,000 cycles learning in one trial. We carry out this trial 5 times and evaluate mean accuracy by cross validation for two types of genuine data. The average discrimination accuracy of this experiment are 97.0 % and 95.9%. In addition, the false rejection rates are 9.6% and 19.2%. The false acceptance rate are 0.8% and 0.1%.

This paper proposes a method to input characters and/or character string in an augmented reality using gesture motion. The proposed method detects the region of character string using gesture motion. It consists of five phases; template generation, skin detection, hand region detection, gesture motion extraction and designation of character string region. The template image consists of two fingers because a gesture is to take hold the tips of the first and second fingers. In the skin detection, we extract the skin color on the basis of values in saturation by using threshold processing. The hand region is detected by calculating areas and detecting the area with the maximum value as a hand. The gesture motion is extracted using template matching. In order to show the effectiveness of the proposed method, we conduct experiments.

Recent years, biological signals have attracted much attention as a tool of human interface. Electromyogram (EMG) has been used in a variety of situations in particular. We measure EMG of arms or shoulders in many cases. In addition, we often use expensive wet type sensors. However, they are inconvenient and high-cost. On the one hand, there have been few works of personal authentication using EMG. Therefore, in this paper we measure EMG by attaching dry type sensors to wrist, and carry out personal authentication. The conventional method in this paper is divided into three units such as a measuring, a feature extraction, and a discrimination units. We measure EMG signals with eight dry type sensors on the wrist. After that, we identify a motion opening our hands. We use a convolutional neural network (CNN) to learning and authentication. In addition, we try to use a multilayer perceptron for comparison. Experiments are conducted in two patterns. At first, we carry out two-class separation (the subject and the others). The second is multi-class separation in which the number of subjects is 8 people. We collected 40 data for each subject. The average accuracy of two-class separation was 89.4 % by the multilayer perceptron. That was 94.9 % by CNN. On the other hand, the average accuracy of multi-class separation was 41.2 % by the multilayer perceptron. That was 70.3 % by CNN. In addition to the conventional method, the proposed method in this paper preprocesses the data. Large noise was removed using a high path filter. By this preprocessing, identification accuracy (Two-class classification using CNN) improved by 1.5%. The true acceptance rate improved by 7.2%, and the false acceptance rate improved by 0.0067%. In future work, we review input data. We consider not only noise removal but also normalization to eliminate intra-individual difference. Furthermore, we review a layer structure of CNN.

In the field of pattern recognition, principal component analysis (PCA) is one of the most well-known feature extraction methods for reducing the dimensionality of high-dimensional datasets. Simple-PCA (SPCA), which is a faster version of PCA, performs effectively with iterative operated learning. However, SPCA might not be efficient when input data are distributed in a complex manner because it learns without using the class information in the dataset. Thus, SPCA cannot be said to be optimal from the perspective of feature extraction for classification. In this study, we propose a new learning algorithm that uses the class information in the dataset. Eigenvectors spanning the eigenspace of the dataset are produced by calculating the data variations within each class. We present our proposed algorithm and discuss the results of our experiments that used UCI datasets to compare SPCA and our proposed algorithm.

Keyword image retrieval is now widely studied. By using such technologies, we can obtain images with the corresponding keywords easily. In the case of conventional image search systems, we basically search according to file names. However, the file names given are frequently incorrect. To resolve this problem, we propose an automatic keyword addition method for scenery images. In this paper, there are two important points. One is the image segmentation method using the maximum distance algorithm (MDA). The other is automatic keyword addition using the color features of regions. In the image segmentation method, we propose an automatic decision method for the parameters of the MDA. For this purpose, we investigate the relation between the optimal parameters and the features of regions. For the color feature extraction of regions, we propose a genetic algorithm (GA). Moreover, in order to show the effectiveness of the proposed method, we provide simulation examples. The results of simulations demonstrate the effectiveness of keyword addition for scenery images. © 2012 Wiley Periodicals, Inc. Electron Comm Jpn, 95(2): 39–49, 2012; Published online in Wiley Online Library (wileyonlinelibrary.com). DOI 10.1002/ecj.10364

Recently, robot technology has gained popularity because of labor shortage, ability to work for long hours, etc. Remote-controlled robots are especially being put to practical use for several purposes such as conveying and security systems, etc. This is because their safety is higher than autonomous robots in addition to their easier development. However, remote-controlled robots require a user interface that can control the robot actions and show information captured by robot to the users. In this work, we develop a simple interface for mobile robots. In robot actions, movement of the robot itself is most important. To use this interface, the users need only point to the position on a screen where they want the robot to move to. After that, the interface controls the robot movement between the current and the target positions autonomously. Therefore, to move the robot to the selected point, the robot needs to calculate the target's 3D location. To calculate the 3D location, we use motion stereo vision that uses a single camera. Using this method, the interface can be applied to many robot systems. In this paper, we proved the effectiveness of the proposed method by performing experiments in real environments. demand a user interface which receives images from the robot and sends orders from the user to the robot. Currently, various devices and methods are used. However, they have several problems. For example, in a control method which sends commands such as FOWARD, ROTATION and STOP, it cannot control the minute movement of the robot. Additionally, another method using game controllers is difficult for new game users. A method using display of the images sent from the robot, directly displaying only the images is common. However, in this method, it is difficult for the user to get the depth of the objects. Moreover, if the distance to the object is too short, the user cannot see it. To solve the problem, an interface that shows the information of range sensors together with the image and one that shows overhead view from another camera mounted above are proposed. However, these methods are expensive and lack versatility.

This paper introduces a novel statistical model for estimating the intraindividual difference in left prefrontal cortex electroencephalogram (EEG) activity, and a method for evaluating the proposed model. It is known that EEGs contain individual characteristics. However, extraction of these individual characteristics has not been reported. The analyzed frequency components of an EEG can be subdivided into components that contain a significant number of features, and components that do not contain such features. From the viewpoint of these feature differences, we propose a model for extracting the features of an EEG. The model assumes a latent structure and employs factor analysis, treating the model error as personal error. We consider the first factor loading, which is calculated by eigenvalue decomposition, as the EEG feature. Furthermore, we use a k-nearest neighbor (kNN) algorithm for evaluating the proposed model and the extracted EEG features. In general, the distance metric used is the Euclidean distance. It is possible that the distance metric used depends on the characteristics of the extracted EEG features and on the subject. Therefore, depending on the subject, we use one of three distance metrics: the Euclidean distance, the cosine distance, or the coefficient of correlation. Finally, in order to show the effectiveness of the proposed model, we present the results of an experiment using real EEG data. © 2011 Wiley Periodicals, Inc. Electron Comm Jpn, 94(5): 9–16, 2011; Published online in Wiley Online Library (wileyonlinelibrary.com). DOI 10.1002/ecj.10326

Recently, the automation of the age estimation technique is hoped for in various fields. Therefore, we propose an apparent-age estimation system using empirical mode decomposition (EMD). Conventional study reported that the time-frequency features are important for age estimation. However, these cannot necessarily extract the time-frequency feature in detail, because the classical technique that have a relationship of trade-off between the time resolution and the frequency resolution are used. On the other hand, the EMD is the novel time-frequency analysis technique that do not have the relationship of trade-off between the time resolution and the frequency resolution. The EMD gives a time-frequency analysis decomposing a signal into several intrinsic mode functions (IMFs). The IMF together with their Hilbert transforms are called the Hilbert–Huang spectrum, which leads to instantaneous frequency and amplitude. We use these features effectively for extracting human's age perception. We estimate the age by a neural network that learns pairs of face image and the Hilbert–Huang spectrum. Furthermore, we compress the data for neural network by using the simple principal component analysis (SPCA). In order to show the effectiveness of the proposed method, computer simulations are done by the actual human data.

In this paper, we propose a fully automatic real time one image face gesture simulation using image morphing. Given a single image of a subject, we create several facial expressions of the face by morphing the image based on prior information stored in a data bank. The process involves the automatic detection of the control points both on the target image and the source data. The source data is a string of frames containing the desired facial expressions. A face detection neural network and a lips contour detector using edges and the SNAKES algorithm are employed to detect the face position and features. Five control points and the lips contour, for both the source and target images, are then extracted based on the facial features. Triangulation method is then used to match and warp the source image to the target image using the control points. In this experiment, using one expressionless face portrait, we create an animation to make it appear like the subject is pronouncing the five Japanese vowels. The final results shows the effectiveness of our method.

The purpose of this study is to develop an interactive interface using eye movement to operate a welfare apparatus, such as a feeding device for an orthopedically-impaired person. A part of the purpose is to eliminate special calibration and re-calibration during the operation. Originalities of this proposed system are as follows: eye sensing with evolutionary processing and interactive operation screen based on some physiological knowledge. The proposed system uses a non-contact type interface by the eye movement. An iris is tracked and eye movement is measured by the evolutionary eye sensing (EES) method. An operation screen is divided into 9 areas, which has visual stimulation using the physiological knowledge. A user can select these areas by eye movement and decide by eye fixation. The effectiveness of the proposed system is evaluated by comparison experiments with 20 subjects. The results indicate that the proposed system is easy to use for first-timer, who can become proficient in operation just after a few exercises.

Over the last quarter century, research in biometric systems has developed at a breathtaking pace and what started with the focus on the fingerprint has now expanded to include face, voice, iris, and behavioral characteristics such as gait. Palmprint is one of the most recent additions, and is currently the subject of great research interest due to its inherent uniqueness, stability, user-friendliness and ease of acquisition. This paper describes an effective and procedurally simple method of palmprint feature extraction specifically for palmprint recognition, although verification experiments are also conducted. This method takes advantage of the correspondences that exist between prominent palmprint features or objects in the spatial domain with those in the frequency or Fourier domain. Multi-dimensional feature vectors are formed by extracting a GA-optimized set of points from the 2-D Fourier spectrum of the palmprint images. The feature vectors are then used for palmprint recognition, before and after dimensionality reduction via the Karhunen-Loeve Transform (KLT). Experiments performed using palmprint images from the ‘PolyU Palmprint Database’ indicate that using a compact set of DFT coefficients, combined with KLT and data preprocessing, produces a recognition accuracy of more than 98% and can provide a fast and effective technique for personal identification.

This paper presents a new algorithm for incremental learning, which is named Incremental Simple-PCA. This algorithm adds an incremental learning function to the Simple-PCA that is an approximation algorithm of the principal component analysis where an eigenvector can be calculated by a simple repeated calculation. Using the proposed algorithm, it is possible to update the eigenvector faster by using incremental data. We carry out computer simulations on personal authentication that uses face images and wrist motion recognition that uses wrist EMG by incremental learning to verify the effectiveness of this algorithm. These results were compared with the results of Incremental PCA that introduced incremental learning function to the conventional PCA.

This paper presents a method for gender and age estimation which is robust to changing facial pose. We propose a feature point detection method, called the adapted retinal sampling method (ARSM), and a feature extraction method. A basic concept of the ARSM is to add knowledge about the facial structure to the retinal sampling method. In this method, feature points are detected on the basis of seven points corresponding to facial organs from a facial image. The reason why we used seven points as the basis of feature point detection is that facial organs are conspicuous in the facial region, and are comparatively easy to extract. As features robust to changing facial pose, skin texture, hue, and the Gabor jet are used for gender and age estimation. For classification of gender and estimation of age, we use a multilayered neural network. We also examine the left– right symmetry of faces in connection with gender and age estimation by the proposed method. © 2008 Wiley Periodicals, Inc. Electron Comm Jpn, 91(7): 32– 40, 2008; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecj.10125

In this paper, real-time tracking and information acquisition of lip region of a person during speech in active video scenes are addressed. The lip is deformed by speech, and the size and orientation are changed by camera motion. The difficulty is mainly due to change of these appearances of the lip. We use template matching with a genetic algorithm to overcome these problems. A high speed and accurate tracking scheme using Evolutionary Video Processing is proposed. Usually, a genetic algorithm is unsuitable for a real-time processing, however, we achieved real-time processing. In our simulations, a tracking accuracy of 91.6% and an average processing time of 26.0 milliseconds per frame are achieved.

Recently, security cameras have been installed at a high rate in places where there are extensive grounds and many humans gather. The number of installed security cameras has been increasing year by year. The main reason is security enhancement including the prevention of incidences of terrorism. Therefore, we propose a method which detects abandoned objects on online by using pan-tilt camera. Above all, we improve problems of the previous method which is based on ST-Patch features and human detection. We make extended ST-Patch features for solving the problem of ST-Patch features. We improve human detection by using deep learning which is based on a convolutional neural network. We conducted preliminary experiments to verify a method of pooling, and then we decided to use Max pooling because its detection accuracy is better than that of Ave pooling. We conducted experiments in five situations to verify usefulness of the proposed method. If the proposed method finds an abandoned object, it saves the object image. We define the abandoned object as an object which human does not subsist near. We could detect the abandoned object in each situation. However, we conducted experiments of the proposed method only in a room. We need to conduct experiments in a wide area to find new problem.

In this paper, high-speed lips tracking and data acquisition of a talking person in natural scenes are presented. Our approach is based on the Evolutionary Video Processing. This method has a trade-off between accuracy and a processing time. To solve this problem, in this paper, we proposed Evolutionary Video Processing with automatic SD-Control. In our simulations, the effectiveness of the proposed method is verified by a comparison experiment. The proposed method improves the performance, speed and accuracy, from 68.4% to 86.2%. Furthermore, it is evaluated that our proposed method can continue to chase the lips region even in such a case. It is demonstrated that the lips region detection and tracking at high speed and with high accuracy is possible, with acquisition of its numerical geometric change information.

This paper presents a new algorithm for feature generation, which is derived based on geometrical interpretation of the fisher linear discriminant analysis (FLDA). This algorithm (Simple-FLDA) is an approximation algorithm that calculates eigenvectors sequentially by an easy iterative calculation by expressing the maximization of variance between classes and minimization of variance in each class without the use of matrix calculation. We carry out computer simulations about recognition of wrist motion patterns by EMG measured from wrist and personal authentications that use face images to verify the effectiveness of this technique. The result was compared with the result of principal component analysis (Simple-PCA).

In this paper, we propose a recognition system of wrist operation by focusing on ElectroMyoGram (EMG) that is the living body signal generated with movement of a subject. In previous research, we only performed pattern recognition by Neural Network (NN) and Fast Fourier Transform (FFT). In contrast, in proposal research, we try to improve recognition accuracy and reduce learning-time of system by combining Multi Discriminant Analysis (MDA) and gradual Principal Component Analysis (PCA) based on the PCA result of EMG data. From results of computer simulation, it is shown that our approach is effective for improvement in recognition accuracy and speed.