Previous reports have shown the validity of using thermal images taken by UAVs, or aerial drones, for solar panel fault detection. In order to use AI algorithms for automatic fault detection, it is necessary to convert the thermal video to still images. When converting the video to still images, if the images can be stitched together to create one large image, it will be easier to determine the exact location of the faulty panel. But since the glass surface of the solar panel is smooth and flat, light is reflected off the surface of the glass and the temperature reading of the thermal camera changes depending on the angle of incidence. Since standard stitching algorithms do not account for images which change color depending on angle, the connected thermal image does not show temperature readings correctly. In this research, a stitching algorithm to solve the above problem is proposed. By using this algorithm, a preprocessing method for preparing a thermal image file suited for AI fault detection is proposed.

Due to the expand of COVID-19, remote job interviews have increased, but many job seeking students feel anxious about them. In this study, an experiment was conducted in which a participant played a role of an interviewer of a remote job interview under some conditions. The participant’s eye gaze was recorded and his or her impression was estimated using SD method. It is revealed that looking at camera is important for an applicant, since the more he or she looks at the camera, the more the interviewer looks at his or her eyes and values his or her personality and trustworthiness.

Anomalies in industrial images can be categorized into logical anomalies and structural anomalies. Logical anomalies refer to irregularities such as object deficiency, excess, or misplacement, while structural anomalies indicate impurities such as dirt, scratches, or foreign matter inclusion. Conventional anomaly detection by normalizing flow efficiently detects structural anomalies by learning the local feature distribution of images. However, these existing methods have struggled with the detection of logical anomalies. In this study, to address this problem, we introduced a self-attention mechanism into the normalizing flow to capture global features during variable transformation. Our proposed method was evaluated and compared with a baseline normalizing flow using conventional convolution layers on the MVTecLOCO dataset. Unexpectedly, we observe that the use of the self-attention mechanism does not improve performance.

In semiconductor manufacturing, circuits on the silicon wafers are inspected in various ways, and the distribution of inspection results is obtained as a wafer map. Since the obtained patterns of wafer maps depend on manufacturing error(s), thus the classification of wafer maps and identification of their causes are essential from the viewpoint of production control. There are many studies on wafer map classification, and methods using Convolutional Neural Networks (CNNs) have been proposed. However, these conventional methods assume only defect patterns that have occurred in the past, and these methods cannot detect unknown patterns, which are unexpectedly occurred and obtained in the actual manufacturing sites. Thus, detecting unknown patterns and classifying known patterns with high accuracy is essential to improve the current yield analysis. In this study, the authors proposed a classification system using plural binary classifiers and investigated a feature extraction for the method. In the method, binary classifiers specialized for each known defect pattern were used, and these were concatenated based on each precision value. We conducted evaluation experiments using WM811K dataset. The results showed that the classification accuracy was 77.1% for known defective patterns, and 30.3% of unknown patterns were picked up with the proposed method.

High strength bolts are used to combine the main parts of steel bridges. In the tightening management, inspectors visually check all the bolts after the final tightening step based on each bolt’s marking status. We have developed a system for automatically detecting the tightening states of high strength bolts in order to prevent inspection omissions, using images captured by a mixed reality device. In this system, images of high strength bolts are extracted from the captured image using YOLO, a deep learning object detection algorithm, and the distance from the normal state is detected using deep metric learning. We have applied a Laplacian filter to calculate the blurriness to exclude excessively blurred images, so that the detection will stay stable even during movement. In operational tests using real steel girders, we confirmed that the system accurately detected the tightening states and displayed the judgment results on the window of a mixed reality device in near real time even while moving.

Recently, the Cabinet Office is advocating the concept of ”Super City” that realize new lifestyles and businesses for fundamentally changing the state of society through AI. The authors have established a research unit of ”Health Smart Network” at Research Institute for Socionetwork Strategies (RISS), Kansai University as a research base to meet these social demands. Specifically, the purpose of this research unit is to make health promotion services smarted with ”eHealth + AI”, create new social communication, and contribute to extending healthy life expectancy and sustaining a healthy life. In this presentation, we outline the concept of Kansai University’s ”Health Smart Network” and discuss the progress and results of various research on the theme of smart health promotion.

Recent years, the number of people with low core body temperature tends to increase, and one of the causes is said to be muscle weakness due to lack of exercise. In addition, it has been pointed out that as the core body temperature decreases, the body's immunity also decreases, making it easier to get sick. On the other hand, with the recent development of machine learning, it is expected to improve the efficiency of work by deep learning in the medical field. Therefore, using machine learning, we proposed a method to visualize physical condition, aiming to predict various biological information only by thermography. In this study, we grasped the relationship between thermographic images and biometric information in step exercise, made the LSTM model and VGG16 model learn the collected data, and verified the prediction accuracy of biometric information. In the prediction experiment, thermographic images and biological information were acquired at three tempos of 100, 120, and 140. Furthermore, we evaluated the health index using the prediction data obtained in this experiment, and verified the error with the actual value.

We have proposed pdi-BoostingG that generates virtual data as a kind of ensemble learning. In this method, region G is set to generate additional fuzzy rules around misclassified data, and virtual data is generated within this region G. Generated virtual data and added fuzzy rules are inherited among layers of ensemble learning. In this paper, we propose a method to generate virtual data with a normal distribution around the internal division points of different classes of misclassified data, with directivity to the central position of the class. Since generated virtual data is inherited between layers together with fuzzy rules, deep inference of fuzzy rules is realized. We discuss here how to generate virtual data and discuss the features of the proposed method from the results of the discrimination rate using numerical examples.

Locomotive syndrome is the decline in mobility function due to aging-related deterioration of the locomotor system, which increases the risk of falling. This study aims to analyze and quantify behavioral components associated with falling risk by measuring daily walking motions. We propose a walking motion analysis method based on the Gaussian Process Dynamics Model (GPDM) to detect changes in body usage. This method compresses gait motion data into a low-dimensional latent variable model using GPDM and visualizes the differences in body movements caused by behavioral conditions through the model’s output.

In this study, we proposed a harmony generation system to promote motivation for group exercise. The system generates sound based on the user’s movement and enables collaborative music performance by overlaying sounds from multiple individuals. Furthermore, the system supports the creation of harmony by correcting the discrepancies in the timing of sound generation between users. In the exper- iment, the effectiveness of the system was verified by having the participants perform various patterns of exercise with and without sound generation and correction, as well as by changing the degree of correction. The results suggest that engaging in exercise using this system can enhance the user’s motivation for physical activity.

The authors are conducting research on the use of smartphones to support the safe walking of the visually impaired persons. Currently, we are developing an application that specializes in functions to prevent accidents involving falls on the station platforms. To evaluate the effectiveness of this application, it is necessary to confirm that it works correctly in crowded places, so the degree of such crowdedness must also be quantitatively determined to quantitatively demonstrate its usefulness. This paper discusses quantitative measures of the crowdedness using a system that measures the location of people, based on experiment data.

Since there have been accidents repeatedly involving the visually impaired persons falling from station platforms, this study tried to develop a smartphone app specifically for use at railway stations. By acquiring distance information with a LiDAR implemented in a smartphone, it will be easier to detect the height difference between station platforms and rail tracks, and the detection accuracy is expected to be improved. This paper shows experimental results using a PC instead of a smartphone to evaluate only the effectiveness of the proposed method algorithm.

Many optimization tasks in the real world can be regarded as Multiobjective Optimization Problems (MOPs) that have multiple objectives to be optimized. Due to trade-off relationships among objectives, MOPs usually have the Pareto-optimal solution set (PS) instead of a single optimal solution. In addition, there is the Pareto-optimal front (PF) which is the image of the PS in the objective space. Decomposition-based Multiobjective Evolutionary Algorithms (MOEAs) are one of the most popular categories of algorithms for MOPs. In the previous study, we introduced CIM-based Adaptive resonance theory (CA), a topological clustering algorithm, into a decomposition-based MOEA to realize the adaptive decomposition according to the PF shape and proposed RVEA-CA. Although several studies show that adaptive decomposition-based MOEAs, including RVEA-CA, have high search performance on MOPs with a large number of objectives and high versatility on the various PF shapes, the effect of adaptive decomposition on constrained MOPs has not yet been investigated, to our knowledge. Hence, this paper introduces a constraint-handling method into RVEA-CA and investigates the search performance on constrained MOPs. The computational experiments showed that the proposed method has search performance equal to or better than those of four state-of-the-art constrained MOEAs and discussed the effectiveness of the adaptive decomposition on constrained MOPs.

With the growth of video streaming and virtual reality, there is a need for services that better match people’s emotional responses. A model can be created to estimate emotions while watching videos and provide media recommendations. EEG signals can be used as input features to estimate latent emotions that are less apparent in facial expressions and gestures. A Multi Layer Perceptron (MLP) can be used for the estimation model, reducing learning time compared to deep learning and gaining generalization ability. The aim is to construct an estimation model using MLP and Power Spectrum Density (PSD) of EEG signals as input features. The average classification performance for 32 individuals was 74.3% for valence and 74.0% for arousal, comparable to a CNN trained using the same input features. We believe that the experimental results show that large networks and convolutional layers are not necessary for EEG-based emotion estimation.

In this study, we propose an improvement to the direct mating method, a constraint handling approach for multi-objective evolutionary algorithm with combining the direct mating and the local mating. The direct mating can maintain the information for the improvement of the design, and the local mating selects another parent from the feasible solution space around the initially selected parent. The direct mating method selects the other parent along the optimal direction in the objective space after the first parent is selected, even if it is infeasible. We evaluate the proposed method on three mathematical problems with unique Pareto fronts and one real-world application. We use the generation histories of the averages and standard deviations of the hypervolumes as the performance evaluation criteria. Our investigation results show that the proposed method can solve constraint multi-objective problems better than existing methods while maintaining high diversity.

The evolutionary computation competition in 2022 introduced an estimation task of traffic volume of pedestrians by using simulation. This task requires participants to estimate the departure time and the number of people of three types of pedestrians (slow, busy, guided), respectively. In this report, we propose to model traffic volume as a mixture distribution and to optimize the center point, width, strength, distribution type, and the width of Gaussian filter by using evolutionary computation algorithms. For one of the single objective optimization problems, our methods came second, which shows the effectiveness of our proposed method.

Various studies have been conducted thus far on robot emotional expression. Our previous research demonstrated the potential for variations in robot emotional expression based on the color of their clothing. Consequently, in this paper, we report the results of an investigation into the effects of emotional expression when using electroluminescent sheets that display coloration, as a material for robot clothing.

Recently, programming education has become compulsory in elementary and junior high schools. Among these, board games that enable students to learn ”programming thinking” are being used at various educational institutions. Board games have the advantage that beginning programmers can learn ”programming thinking” such as sequential execution and iterative processing without stress by playing while having fun. However, board games require the presence of a partner, and it is not always possible to have a partner to play with at home. Therefore, this study develops a partner-type robot that can learn programming educational board games together. The experiment verifies the learning effect on the learner by comparing a board game between two people and a game with our robot.

In recent years, the proportion of children with developmental disorders in regular classrooms has been increasing, leading to a growing demand for intelligence test. Since only clinical psychologists can administer this test, there is an increasing need for their services. Currently, training for intelligence test involves demonstrations with children and role-playing among adults pretending to be children. However, it is difficult to conduct sufficient training due to the challenge of performing demonstrations with children. Therefore, this study develop a child type robot (the proposed robot) capable of training clinical psychologists in intelligence test. This paper examines whether clinical psychologists can conduct training for intelligence test using the proposed robot. In the experiment, experienced clinical psychologists will be asked to administer intelligence test to 3-year-old children using the proposed robot, and their ability to adapt to the training will be investigated through interviews.

In this study, we propose a robotic model that creates a sense of distance by using a conversational strategy based on politeness theory. The ”sense of distance” in dialogue refers to the perceived emotional or social distance between the participants in a conversation. When it comes to human-robot interaction, robots are typically designed to have a consistent conversational style or sense of distance, regardless of the user’s level of intimacy or familiarity. However, this uniformity can result in a lack of human-like behavior and contribute to the sense of artificialness. Recently, the number of robots communicating with humans has been increasing, and further development is expected by achieving familiarity and human-like qualities. Therefore, in this study, we focus on the ”sense of distance” in conversation and attempt to create variations in distance by modifying linguistic behavior. In this experiment, we employed several conversational strategies centered around a simulated two-day tour guide theme and gradually reduced the sense of distance produced by the robot. As a result, approximately 80% of the participants confirmed perceiving a sense of human-likeness and familiarity in the in the way the robot established social dialogue distance.

Previous research has shown that a faster pace in human speech increases the impression of intelligence. Another research has confirmed that there’s a correlation between the robot’s intelligence and the frequency (support acceptability) with which advice from that robot is heeded. This paper reports the results of a study investigating how the speech speed of a driver assistance robot influences the support acceptability.

Recently, educational support robots have attracted much attention. In conventional educational support robots, the number of questions to be solved by the learner is fixed. However, in order to encourage learners to learn spontaneously and to improve the learning effect, it is important for learners to be aware that they themselves try to solve a larger number of problems. Therefore, we considered that it would be effective for the robot to suggest that the number of problems to be solved by the learner should be increased by using the politeness theory. Therefore, this study constructs the method which the robot proposes the number of problems based on the politeness theory. The robot is equipped with three different politeness strategies, and by comparing each strategy, we investigate the impression effect that the robot's speech with each strategy makes on the learner.

In recent years, ICT education has been introduced to the educational field, and research and development of educational support robots have been attracting attention. There are teacher-type robots that instruct the learner about the contents of learning, just like a teacher. Conventional teacher-type robots provide learning support by pressing a button, but this method has been reported to cause excessive demand for support from the learner. To solve this problem, we proposed a method to estimate the learner’s state of perplexion based on the learner’s facial expressions through deep learning. Furthermore, we constructed an apprenticeship promotion model by combining a behavioral model that provides learning support based on cognitive apprenticeship theory and the perplexion estimation method. The experiments confirmed that the robot equipped with the apprenticeship promotion model provides the same level of learning effect to university students as a conventional robot. In this paper, we design the room which our robot learn with a junior high school and investigated whether junior high school students use the robot for learning.

Premature newborns are usually admitted to the neonatal intensive care unit (NICU) for several weeks to receive advanced medical management. However, we need to investigate the NICU environment because the light and noise-emitting monitoring devices and medical equipment can adversely affect the circadian rhythm, which is the sleep-wake cycle of newborns. There are methods and devices available to measure the sleep-wake state of newborns, but they can be burdensome to newborns and nurses. Therefore, this study proposes a non-contact, automatic newborn wakefulness state classification method. The proposed method classifies sleep-wake states using 3DCNN from the entire body and the face region video. Experimental results showed that, by combining the results of the two videos, the classification performance was improved from our previous research using OpticalFlow and SVM.

Constipation has multiple symptoms and pathologies, and the treatment depends on the cause of constipation. Quantification of the amount and location of gas and stool in radiographic images is effective to select the appropriate treatment. We are developing a segmentation method of gas and stool regions in radiographic images using U-Net. The purpose of this paper is to investigate the effectiveness of pre-learning and backbone in U-Net. The gas volume score (GVS) is an index for quantitatively evaluating the quantity of intestinal gas. As well as GVS, we propose a stool volume score (SVS), and the joint volume score (JVS) which quantify the combined region. Experiments were conducted under two conditions: with/without a backbone, and with/without pre-training. The experimental results showed that the proposed method with vgg16 backbone and fine-tuning achieved a correlation coefficient of 0.901 for GVS, 0.618 for SVS, and 0.437 for JVS. DICE coefficients were 0.669 for the gas region, 0.523 for the stool region, and 0.646 for the combined gas and stool region. The extraction accuracy was best when both pre-training and backbone were used, indicating the usefulness of pre-training and VGG16 backbone in this method.

Segmentation of major brain tissues from 3D medical images can contribute to improving diagnostic quality and reducing workload. This study aims to explore the proper structure to segment brain regions from MRI volumes. The dataset used was the preprocessed IXI dataset, and segmentation was performed on 46 regions from the head MRI images. Experimental results show that it is important to extract features in the first stage, when the resolution is large.

Circulating Tumor Cells (CTCs) are cancer cells that have detached from the primary tumor or metastatic sites and circulate in the bloodstream. CTCs are considered to be factors in cancer progression and metastasis, and their detection enables early cancer detection, prediction of disease progression, and monitoring of treatment effectiveness. Electro-rotation of cells is a phenomenon in which cells undergo rotational motion due to an externally applied rotational electric field. It is suggested that by measuring the rotational speed, which varies due to the electrical characteristics of cells, it is possible to distinguish between CTCs and normal cells. In this study, we aimed to automatically measure the rotation speed of living cells by electro-rotation using video processing technology in order to realize the discrimination between CTCs and normal cells. In automatic measurement, it is difficult to measure the rotation speed from a feature point on the shape or a fixed reference point, which is generally done, because the rotation direction is not necessarily horizontal to the camera and orbital rotation is taking place. Therefore, we attempted to solve the aforementioned problems by sequentially updating feature and reference points.

This study aims to establish a method for analyzing patients’ gait without special equipment (such as a motion capture system). Routine and quantitative gait analysis is crucial for diseases with gait disorders (e.g., Cerebral Palsy). Unfortunately, gait analysis requires expensive motion capture systems and highly trained personnel. This paper addressed the challenge of estimating the deviation of patients’ gait from standard case, employing the two-dimensional positions of joints estimated by off-the-shelf human pose estimation with Auto-Encoder (AE). The authors calculated the correlation coefficient between the reconstruction error and GDI to evaluate the method and obtained the result that the correlation coefficient r is -0.395. The experimental results suggest the proposed method could capture the patients’ gait deviation from normality.

The hand movements during precision work are an important factor to evaluate ability and quality. However, it is not easy to evaluate worker’s skill objectively and teach the skill to beginners, because there are individual differences in hand movements. In order to improve the learning efficiency of precision work, we develop an evaluation system to evaluate worker’s skill. The system is composed of hand tracking and fuzzy inference. We employ MediaPipe as a hand tracking tool. The worker’s skills are evaluated by fuzzy inference based on hand movements data obtained by MediaPipe. In This study, we focus on soldering as a precision operation.

Arteriovenous crossing phenomenon is a change in the shape of veins such as stenosis and distortion at the intersection of arteries and veins in the retina, and is one of the important indicators of arteriosclerosis. In estimating the degree of arteriosclerosis based on the shape of veins, the accuracy of contour line extraction of the blood vessel wall is important. In this study, to extract the shape of the vein wall near the intersection, which has a complex shape, without transferring to the artery wall, we investigated a method of blood vessel wall tracking in which the search direction is sequentially updated with the direction of the blood vessel centerline as the reference direction. The effectiveness of the proposed method was verified by experiments using actual fundus images.

Gaussian Process Sequential Regression Models (GPSRM) is a method to obtain nonlinear cluster structures without requiring the number of clusters. We propose a Markov chain Monte Carlo-based hyperparameter optimization for GPSRM. To evaluate the performance of the proposed method, numerical experiments were conducted on artificial datasets with nonlinear cluster structures. The results show that the proposed method performs better than existing methods at the maximum value of ARI.

Estimating the appropriate number of clusters is a problem in clustering research. To solve this problem, automatic cluster number estimation methods have been proposed. The cluster validity index and BIC methods have also been studied to estimate the appropriate number of clusters for hierarchical clustering. In this study, we examined the effectiveness of the automatic cluster number estimation method proposed in previous studies and our proposed method. Our proposed method uses the trace of the covariance matrix to calculate the merging level. DBSCAN and X-means were used as the comparison method. Numerical experiments suggest that the proposed method shows comparable or better results than comparative methods.

SCAN is a network data clustering method that represents the structure of nodes by their neighborhood sets and extracts clusters, hubs, and noise using a similarity called structural similarity. Since structural similarity is a set similarity, the Jaccard and Dice coefficients, which are also set similarities, can be used to express similarity based on node neighborhood sets. In this paper, we examine the characteristics of clustering results based on the Jaccard coefficient and other similarities of each set. First, clustering is performed using the Jaccard coefficient and other similarities in SCAN, and the results are evaluated by ARI and Modularity. Next, we visualize the similarity of the clustering results using the multidimensional scaling. From the experimental results, it was confirmed that the Jaccard coefficient and other set similarities can extract clusters that are difficult with Modularity, and that SCAN using the Simpson coefficient in particular tends to show a high ARI.

Noise clustering is a clustering method that classifies outliers as noise clusters and is known as a method to reducing the influence of outliers. Local Outlier Factor is an anomaly detection method that quantifies the degree of data outliers based on the density with neighboring data. In this study, we propose a new noise clustering method that introduces the Local Outlier Factor. Numerical experiments show that the proposed method is more robust to outliers than existing methods for artificial and real datasets.

In clustering-based collaborative filtering (CF), clusters of users with similar preference patterns are extracted, and items with high preferences within the cluster are recommended. Since data in CF tasks contain uncertainties arising from human sensibilities, represented as co-occurrence relationships between users and items, approaches such as rough clustering and co-clustering can be effective. Thus, rough co-clustering induced by multinomial mixture models (RCCMM) and its application to CF (RCCMM-CF) have been proposed. However, RCCMM has a problem in that it does not consider the granularity, an important viewpoint in rough set theory. In this study, we propose a CF approach based on rough set-based co-clustering induced by multinomial mixture models (RSCCMM) considering granularity. Furthermore, we verify the recommendation performance of the proposed method through numerical experiments using real-world datasets.

Privacy preserving data clustering is a useful method for extractingintrinsic cluster structures from distributed databases keeping personal privacy. In this research, a novel model of performing Fuzzy c-Lines clustering with horizontally distributed data is proposed, where federated learning is achieved by sharing gradient information estimatedin each client. The proposed model is an extension of the Fuzzy c-Means-type federated learning model proposed by Pedrycz to linear clustering with least square criterion.

Intra-Cluster COntrolled Edge-sized Clustering Based on Optimization for Network data (ICCOECBON) is a clustering method for network data. Since ICCOECBON is based on k-medoids, its clustering results are dependent on initial values. In this study, we propose ICCOECBON++ as a method that introduces the initial value determination method of k-means++ into ICCOECBON. Experimental comparison with existing methods was conducted on four benchmark datasets. From the experiments, it is confirmed that the proposed method improves the initial value dependence of ICCOECBON.

In recent years, many studies have been conducted to elucidate the information processing mechanism of the human brain using deep learning. In this study, to investigate the relationship between visual and language information in the human brain, we conducted encoding modeling to predict brain activity based on features of videos extracted from the hidden layers of deep learning models, convolutional neural networks for visual information, Transfomer for language information, and a multimodal processing model, CLIP. Then, we analyze brain states predicted by the encoding models to investigate the hierarchical properties of the localization and representational content of visual and language information in the cerebral cortex. As a result, we found that the similarities in predictable brain regions are similar to those in spatial information representation content, but not similar to those in temporal information representation content.

Deep learning highly effective in pattern recognition across various fields, but obtaining sufficient training data in the life sciences fields can be challenging due to the nature of biological data. In this study, we classified electrical activity pattern evoked by 2 different stimuli to a cultured neuronal network, by converting the time-ordered sequence of instantaneous spatial patterns into an image. These patterns were used as training and validation data for transfer learning with pretrained model, VGG-16. Instantaneous spatial patterns were generated by dividing the continuous evoked responses into 1 ms time windows, with the pixel brightness representing the firing rates for 10 evoked sweeps. Two types of images were provided to VGG- 16: “ Spatial Information Priority Neural Activity Pattern ”(SIP-NAP) images, which preserve electrode arrangement within the same time window and rearranged multiple images, “ Time Information Priority Neural Activity Pattern ” (TIP-NAP) images, which rearrange multiple images to ensure that adjacent brightness values (firing rates) of the same electrode are in different time windows. Classification accuracy evaluated for each type of image, with successful classification achieved at over 90% accuracy and rapid convergence in TIP-NAP. These results suggest the important role of temporal information flow, or the “ stream ”, plays an important role in representing information in a neuronal network.

Understanding of human flexible cooperative behavior is important for making social agents and robots, and also construction of cooperative models and simulations are significant for constructive understanding of human group behavior. We have analyzed human group behavior including intention inference in groups and modeled their action selection process using cooperative pattern task so far. As a result, while some methods of intention inferencing and differences of roles in the groups have been revealed, subjects’ behaviors are not enough to explain. Therefore to extend our model, we implement leading behaviors for explicit expression of their roles and intention in the groups observed in subjects experiments in the pattern task.

In this study, we conducted an experiment to evaluate the degree to which people can recognize guide lights and emergency exit lights placed in commercial facilities under conditions in which they are actively aware of evacuation. The eye movements of the subjects will be measured during the experiment so that their sign-seeking behavior can be analyzed. Considering the difficulty of conducting the experiment in a commercial facility, the experiment will be conducted using a virtual reality system in which 3D images taken at the places are reproduced in a virtual space and presented to the subject on a head-mounted display. From the experimental results, the visibility of evacuation signs and the eye movements used to search for them are discussed.

This paper discusses robust stabilization for fuzzy systems with uncertainties in Takagi-Sugano fuzzy model. In actual system control, model errors always exist in system parameters obtained through system identification. Therefore, in order to improve control quality of a fuzzy controller, it is necessary to consider a control design for the fuzzy systems with the uncertainties of the system parameters. In this paper, we propose a new controller based on Lyapunov function that includes integral functions of membership functions for fuzzy system with uncertainties.

This paper presents a new approach to complicated path generation and stabilization control for an unmanned aerial vehicle (UAV). UAVs are complicated, nonlinear, and unstable systems. Even if the control system is designed by considering the longitudinal and lateral dynamics separately, the coupling between them cannot be ignored. Therefore, it is extremely difficult to design a control system that can follow complicated paths without tracking errors. This paper proposes a control-theoretic approach to complicated path generation and stabilization control, and demonstrates the effectiveness of the proposed approach through flight experiments on a real UAV. The experimental results show that the proposed approach provides excellent control results for a number of complicated paths.

A method of a fusion of fuzzy inference and policy gradient reinforcement learning has been proposed that directly learns, as maximizes the expected value of the reward per episode, parameters in a policy function represented by fuzzy rules with weights and membership functions. A study has applied this method to a task of speed control of an automobile and has obtained correct policies with learned weights of rules, some of which control speed of the automobile appropriately. However, membership functions that quantify fuzzy concepts were designed based on human knowledge. Therefore, in this research, we show the result of experiments that the fusion method can learn the membership functions represented by a layered neural network.

In this study, we proposed a piecewise modeling method using numerical data. The vertex values of hyperrectangular regions were determined using the learning algorithm based on a simplified fuzzy inference model because the piecewise model was represented by a fuzzy if-then rule with singleton consequents. The proposed algorithm can be used to determine the vertex values and positions of segmented regions and with minimum modeling errors. A example was considered to demonstrate the effectiveness of the proposed method using numerical simulations.

This study proposes an evaluation method for fruit tree growth status using automatic annotation that does not require human intervention. The proposed method calculates the normalized difference vegetation index (NDVI) from hyperspectral images taken during the grape growing season, extracts the image values for each small area from the NDVI image, learns the correspondence between the shooting date and time data and the small area by deep learning, and estimates the period for the unlearned small area. The proposed method was applied to hyperspectral images taken in 2021 and 2022, and the growth status was estimated using the model trained on the small areas extracted by equal division. The growth rate was defined as the ratio of the number of small areas that were judged to be the same as the shooting period among all small areas, and the growth comparison showed that the growth progress is likely to be similar until around mid-August. The impact of downy mildew, which occurred frequently from around August 4, 2022, on the growth rate was not suggested, and the possibility that pinching, one of the cultivation management operations, affects the NDVI image was suggested.

In recent years, machine learning technology has made significant advancements, and it has also been utilized in various tasks in natural language processing. One of these tasks is ”aspect-based summarization,” which involves summarizing information based on the important aspect or viewpoint within a text. In this study, we propose a method for automatically extracting evaluation target words and evaluation terms, which are crucial elements related to the ”aspect,” using a deep language model. We utilize the ”chABSA-dataset” created and released by TIS Corporation as part of aspect-based summarization research and examine the effectiveness of the proposed method through experiments.

This study examines whether discriminant analysis or neural networks are more effective in predicting the utterance or no utterance, using lip features as explanatory variables. First, the maximum amplitude and frequency derived from a lip movement wave and the coordinates of the four fixed points in a lip are defined as feature values. As for the coordinates, three cases are set where both x and y coordinates and one of them are used. Second, by applying these feature values to discriminant analysis and neural networks, the utterance or no utterance is predicted. Consequently, it is shown that a neural network in which only the y-coordinate of lips is used as the explanatory variables guarantees high prediction accuracy.

In Japan, where the population is aging, falls are a serious cause of bed-ridden and death among the elderly, and falls need to be detected. However, it is very difficult to collect fall data to train a model to detect falls. The purpose of this study is to generate fall data using a Generative Adversarial Network and to evaluate whether falls can be detected using this data.

This paper presents an approach for localizing basketball players and objects in real-world coordinates from videos. This is achieved by detecting 33 key points spread across the basketball court and using them as a reference to find the real-world coordinates of any pixel in the image that is known to be part of the ground plane. Moreover, a comprehensive synthetic dataset of basketball court images was generated using Blender: Each image was associated with a list of the 33 points’ 2D coordinates in the image. The key points are detected by a model performing two tasks: classification, determining whether a point is visible in the image, and regression, estimating the point’s precise location within the image. While the model was able to achieve more than 80% accuracy in detecting key points in the synthetic images, it was not able to generalize to real-world images, emphasizing the need to label some real-world data to fine-tune the model in the future.