Hiroshi Higashi

Multilinear Discriminant Analysis With Subspace Constraints for Single-Trial Classification of Event-Related Potentials

The classification accuracy of a brain-computer interface (BCI) frequently suffers from ill-posed and overfitting problems. To avoid and alleviate these problems, we propose a method of a multilinear discriminant analysis with constraints to augment parameter reduction, regularization, and additional prior information for event-related potential (ERP)-based BCIs. The method reduces the number of parameters by multilinearization, regularizes the ill-posedness via subspaces that constrain the parameter spaces, and incorporates a brain functional connectivity through the constraints. The experimental results show that the proposed method improved the classification accuracy rates in a single-trial ERP processing.

Dynamic Visual Cues for Differentiating Mirror and Glass

Mirror materials (perfect specular surfaces such as polished metal) and glass materials (transparent and refraction media) are quite commonly encountered in everyday life. The human visual system can discriminate these complex distorted images formed by reflection or transmission of the surrounding environment even though they do not intrinsically possess surface colour. In this study, we determined the cues that aid mirror and glass discrimination. From video analysis, we found that glass objects have more opposite motion components relative to the direction of object rotation. Then, we hypothesised a model developed using motion transparency because motion information is not only present on the front side, but also on the rear side of the object surface in the glass material object. In materials judging experiments, we found that human performance with rotating video stimuli is higher than that with static stimuli (simple images). Subsequently, we compared the developed model derived from motion coherency to human rating performance for transparency and specular reflection. The model sufficiently identified the different materials using dynamic information. These results suggest that the visual system relies on dynamic cues that indicate the difference between mirror and glass.

Spectral-difference enhancing illuminant for improving visual detection of blood vessels

The goal of this study is to develop a functional illuminant which assists medical institution workers who use injection equipment. In this paper, the functional illuminant for supporting people in distinguishing blood vessels from skins was developed. This functional illuminant consists of three types of LEDs. The illuminant with various LED was simulated and the three types of LEDs are selected in such a way that the resulted illuminant emphasizes the color difference between skins and blood vessels. Moreover, we developed a prototype of this functional illuminant and tested it. The test shows that the functional illuminant can support people in distinguishing blood vessels from skins. Furthermore, our experiments suggest that the functional illuminant is robust to individual difference in skin color.

Variation in Event-Related Potentials by State Transitions

The probability of an events occurrence affects event-related potentials (ERPs) on electroencephalograms. The relation between probability and potentials has been discussed by using a quantity called surprise that represents the self-information that humans receive from the event. Previous studies have estimated surprise based on the probability distribution in a stationary state. Our hypothesis is that state transitions also play an important role in the estimation of surprise. In this study, we compare the effects of surprise on the ERPs based on two models that generate an event sequence: a model of a stationary state and a model with state transitions. To compare these effects, we generate the event sequences with Markov chains to avoid a situation that the state transition probability converges with the stationary probability by the accumulation of the event observations. Our trial-by-trial model-based analysis showed that the stationary probability better explains the P3b component and the state transition probability better explains the P3a component. The effect on P3a suggests that the internal model, which is constantly and automatically generated by the human brain to estimate the probability distribution of the events, approximates the model with state transitions because Bayesian surprise, which represents the degree of updating of the internal model, is highly reflected in P3a. The global effect reflected in P3b, however, may not be related to the internal model because P3b depends on the stationary probability distribution. The results suggest that an internal model can represent state transitions and the global effect is generated by a different mechanism than the one for forming the internal model.

Spatial smoothing of canonical correlation analysis for steady state visual evoked potential based brain computer interfaces

Brain computer interface (BCI) is a system for communication between people and computers via brain activity. Steady-state visual evoked potentials (SSVEPs), a brain response observed in EEG, are evoked by flickering stimuli. SSVEP is one of the promising paradigms for BCI. Canonical correlation analysis (CCA) is widely used for EEG signal processing in SSVEP-based BCIs. However, the classification accuracy of CCA with short signal length is low. In order to solve the problem, we propose a regularization which works in such a way that the CCA spatial filter becomes spatially smooth to give robustness in short signal length condition. The spatial filter is designed in a parameter space spanned by a spatially smooth basis which are given by a graph Fourier transform of three dimensional electrode coordinates. We compared the classification accuracy of the proposed regularized CCA with the standard CCA. The result shows that the proposed CCA outperforms the standard CCA in short signal length condition.Brain computer interface (BCI) is a system for communication between people and computers via brain activity. Steady-state visual evoked potentials (SSVEPs), a brain response observed in EEG, are evoked by flickering stimuli. SSVEP is one of the promising paradigms for BCI. Canonical correlation analysis (CCA) is widely used for EEG signal processing in SSVEP-based BCIs. However, the classification accuracy of CCA with short signal length is low. In order to solve the problem, we propose a regularization which works in such a way that the CCA spatial filter becomes spatially smooth to give robustness in short signal length condition. The spatial filter is designed in a parameter space spanned by a spatially smooth basis which are given by a graph Fourier transform of three dimensional electrode coordinates. We compared the classification accuracy of the proposed regularized CCA with the standard CCA. The result shows that the proposed CCA outperforms the standard CCA in short signal length condition.

Smoothing of xDAWN spatial filters for robust extraction of event-related potentials

The xDAWN algorithm is well-known as a method for designing spatial filters to improve signal-to-noise ratio and to reduce the dimension of observed EEG signals. This paper proposes a method for spatially smoothing xDAWN spatial filters to give a robustness against small sample problem. The proposed method gives a subspace constraint to the parameter space of the spatial filters. This subspace is given as a basis of the graph Fourier transform on a graph representing geometric structure of the electrodes. The spatial filters found in the subspace are smooth in the spatial domain. The proposed method is evaluated by experiments with artificial signals and BCI datasets.

A visualization method for hand cleanness using fluorescent spectrum

The goal of this study is to develop a method for the visualization of the distribution of the contamination of the human hands by using fluorescent spectrum. To archive this, we first found the optimal combinations of the excitation and the emission wavelengths that accurately predict the contamination level. This optimization was conducted using the excitation-emission matrices of the contamination components wiped from hands. Then, we visualized the distribution of the contamination on hands by using the spectral imaging of the hands measured with the optimal wavelength combinations. This result suggests that the proposed method can detect and visualize the contaminated areas of the hands.

Optimization of illuminant spectrum for visual detection of foreign substances in jams

Carrying out accurate visual inspection to remove the foreign substances is a crucial issue in food industry. Also in jam factories, the visual inspection by inspectors has been strengthened. However, it is difficult to remove the foreign substances perfectly because they are stained with fruit juice and their color becomes similar to the jam. To overcome the problem, we have proposed the functional illuminant that emphasizes the color difference between jams and foreign substances for the visual inspection. Our previous method for developing the functional illuminant is time consuming to design spectrum. Moreover, this method has low flexibility because it designs an illuminant whose spectrum is produced by a combination of only three types of LEDs, and it has a constraint that the color of the illuminant must be specified before designing. In this study, we propose a new method that formulates the design procedure as an optimization problem solved by NLP (nonlinear programming). The proposed method can combine more than three types of LEDs. Furthermore, our method does not require a constraint for the color of the illuminant. As a result, the color difference by the functional illuminant designed by the proposed method is bigger than the previous method.

Subspace-constrained multilinear discriminant analysis for ERP-based brain computer interface classification

Classification in brain computer interfaces (BCIs) frequently suffers from small sample problem which leads an ill-posed problem and overfitting/overtraining. To avoid and reduce the problems, we propose a multilinear formed linear discriminant analysis with constraints which are derived from other datasets. The proposed method prevents the ill-posed problem by reducing variables and improves robustness by transferring information from other datasets. The experimental results show that the proposed method improves classification accuracy in event-related potential-based BCIs.

Objective assessment and quantification of pearl quality by spectral-spatial features

Quality of pearl is subjectively and visually assessed by experts. The experts assess pearls from various factors such as size, shape, gloss, color, and so on. Among these factors, body color, interference color, glossiness and thickness of nacre of pearl are the most important for the quality of pearls. This study proposes a method that assesses and quantifies each factor of the pearl quality by using optical measurement. The spectral and spatial features were extracted from spectral imaging measurements. Quantitative values for each factor are estimated from the spectral and spatial features. We compared the estimated quantitative values with experts evaluations. The result shows that the proposed method can accurately predict the experts evaluations.