This study proposes a method for modeling the emotional and impression-related attributes of Japanese gait onomatopoeia based on character-level features. Fourteen representative gait-related onomatopoeic expressions were evaluated on six basic emotions and four impression-related dimensions through human annotation. Using these scores as supervised training data, a regression model employing a Character-level Convolutional Neural Network (Char-CNN) was trained. The model demonstrated the capability to predict emotional and impression values not only for known onomatopoeia but also for previously unseen ones. The results suggest that the structural features of onomatopoeic expressions contribute to impression formation and emotional perception.

The phenomenon in which a sound evokes a particular image is referred to as sound symbolism. While numerous studies on sound symbolism have been reported, many aspects of the phenomenon remain unexplained. Our ultimate goal is to elucidate the full picture of sound symbolism. To that end, we aim to uncover its underlying principles through a bottom-up analysis of various real-world examples. In this study, we focused on the names of main characters appearing in two comic book magazines that target different age demographics, selecting them as a case study in which sound symbolism is presumed to be involved. This approach was based on the hypothesis that readers’ age-related impressions influence the phonological structure of the names of the main characters. In the experiment, we performed binary classification of the names of the main characters into two categories: Monthly CoroCoro Comic and Weekly Shonen Jump. The classifier was built using a convolutional neural network (CNN), and feature analysis was conducted using Grad-CAM, a method from the family of explainable AI (XAI) techniques. Based on the experimental results, we hypothesized that the presence of the mora /N/ evokes a CoroCoro-like image. This hypothesis was subsequently validated through a subjective evaluation experiment and statistical testing, both of which supported the hypothesis.

This study investigated the effects of onomatopoeic instructions―commonly used in sports coaching―on one-handed handball grasping actions. To this end, a handball embedded with a force sensing resistor (FSR) was developed. In the experiment, undergraduate and graduate students were instructed to grasp the ball using phrases that each included a specific onomatopoeic expression, such as “Hold it tightly, like ‘Gu’.” The force applied to the FSR on the ball’s outer surface by the thumb during grasping was measured for each type of onomatopoeic instruction. Additionally, participants’ subjective interpretations of each onomatopoeia were collected via questionnaire. The results demonstrated quantitative differences in grasping actions as well as individual variations in the interpretation of onomatopoeia. These findings underscore the importance of quantifying and visualizing individual differences in the interpretation of instructions and physical expression in sports coaching.

In this study, we investigated the effects of onomatopoeia instructions on physical actions in sports coaching. We used a thin-film pressure sensor (Force Sensing Resistor) embedded in a handball. We measured grip pressure data from undergraduate and graduate students under various onomatopoeic instructions. We also investigated subjective interpretations of each onomatopoeia. We applied machine learning-based clustering method to examine differences in grip motion patterns by onomatopoeia type. We explored the relationship between phonological characteristics of onomatopoeia and actual physical movement data. Our results indicate that onomatopoeic instructions influence movement through multiple factors, involving both individual interpretation and physical expression, rather than simple one-to-one correspondence.

In this study, we propose and demonstrate a simple method for fabricating a condenser microphone by attaching iron-on conductive fabric to a commercially available disposable mask. We employ the “Easy-to-Use Wearable Conductive Fabric” developed by Mitsuya Co., Ltd. which allows users to fabricate a microphone quickly and inexpensively using only a household iron, without requiring specialized tools or technical expertise. This conductive fabric offers not only strong adhesion when ironed but also excellent flexibility and breathability, allowing the mask to retain its original comfort and disposability while easily acquiring microphone functionality. Given these characteristics, the proposed method is well-suited for practical use in wearable devices for daily life. The fabricated device functions as a condenser microphone with the conductive fabric and the nonwoven mask serving as electrodes and dielectric, respectively. Sound-induced vibrations cause changes in capacitance, which are detected as voltage variations. This study demonstrates that a fully functional microphone can be realized using only readily available materials and a straightforward fabrication process.

In recent years, gaze input interfaces have been gaining attention as a contactless control method for individuals with physical limitations. In this study, we developed a gaze-based input interface for the remotely operated communication robot “ Marut ”allowing users to control the robot by fixating on specific regions of a PC screen. The system interprets sustained gaze as intentional input, enabling both robot movement and the switching of emotional expressions. Furthermore, we conducted an evaluation experiment to examine whether the emotional expressions used in this study were effectively conveyed by Marut and to assess the impressions they gave to users.

This study proposes a control system for a motorized cart that follows a user via wire manipulation, aiming to achieve natural following behavior with zero-tension perception. In conventional wire-based following systems, the cart is controlled by detecting the extension or retraction of the wire, which inevitably causes the user to perceive tension. In contrast, the proposed system utilizes a camera mounted on the cart to detect the angle and motion of the wire. By implementing a predictive control algorithm that drives the cart before the user perceives tension, the system aims to realize smooth following behavior with zero-tension experience.

This study proposes the construction of an educational program for learning heavy equipment operation and programming in agriculture and disaster prevention using a heavy equipment simulator that can be controlled from a block programming environment. In the proposed method, we developed a heavy equipment simulator that can be controlled from the block programming language Scratch using the game engine Unity, and constructed a tutorial on heavy equipment operation related to agriculture and disaster recovery work. As a result of a hands-on experience using the developed system, participants enjoyed the programming and found the system more convenient and useful in times of disaster.

In recent years, the development of programming thinking skills has been emphasized in school education. In the research of educational support robots, which assist human learning, various robots have been developed to cultivate such skills in children. Our study focuses on a partner-type robot that promotes programming thinking through collaborative learning using board games designed for programming education. Conventional robots have employed the Urge system to generate expressions of emotion that resemble those of humans during question answering, thereby enhancing the sense of joint learning. Moreover, previous studies have suggested that emotional expressions during the process of judging correctness also contribute to a positive impression on the learner. Therefore, this study proposes a partner-type robot capable of expressing emotions both during response generation and during correctness evaluation. An experiment was conducted to compare the impressions formed by university students of four different robots, each exhibiting distinct processes of emotional expression.

In recent years, educational-support robots that assist human learning have attracted increasing attention, and various studies have reported their effectiveness. Efforts are also underway to promote the broader use of these robots through network-based services, including the development of the Robot Service Network Protocol (RSNP), a communication protocol specifically designed for robotic applications. However, when deploying educational support robots over a network, communication delays―such as those caused by receiving questions or instructional content―may cause stress for learners. In network-based services, Quality of Experience (QoE), which reflects the user’s subjective perception, is a critical factor. Despite this, previous studies on educational-support robots have not addressed QoE evaluation. Therefore, this paper proposes a networked educational-support robot system based on RSNP and investigates the impact of content presentation delays on learners. In our experiment, two question display methods―one synchronized with RSNP reception timing and the other using a fixed delay―were randomly presented to participants. The QoE was then evaluated through subjective assessments provided by the learners.

This study tackles the challenge that the eCO2 output of low-cost MOX-based CO2 sensors, estimated from TVOC measurements, is highly sensitive to environmental factors―particularly barometric pressure, along with temperature and humidity―making it unreliable for direct use in indoor occupancy estimation. To address this issue, we propose a regression-based calibration approach that explicitly incorporates barometric pressure as a key input. By co-locating a MOX-type sensor with a high-precision NDIR reference instrument in the same environment and continuously collecting multi-day data on CO2, TVOC, temperature, humidity, and pressure, we train a calibration model. The calibrated eCO2 values, combined with temporal and environmental features, are then used as inputs to an indoor occupancy estimation model. We then input the calibrated eCO2 values together with temporal and environmental features into an indoor occupancy estimation model and demonstrate its effectiveness by comparing performance against a baseline that uses the raw eCO2 output.

In the Shikoku region, it is expected that highway toll policies will promote inter-regional exchange. Therefore, this study aims to describe inter-regional exchanges as a model. The attractiveness of a region is expressed by various indicators of the destination. The costs of inter-regional exchange include the amount of consumption during the stay and generalized travel costs. Expressway tolls are included in generalized travel costs. In order to develop a model of regional exchange across Japan, the difference in generalized travel costs between outbound and return journeys is taken into account. The indices of dwell time and number of visits gave reasonable results.

In the field of social infrastructure, equipment inspections are still predominantly carried out through visual observation. In recent years, image classification techniques based on Convolutional Neural Networks (CNNs) have garnered significant attention for the purpose of automating this process. However, it is often challenging to obtain sufficient data from real-world sites. Therefore, transfer learning using models pretrained on large-scale datasets has become a common approach. In transfer learning, convolutional layers are typically frozen, which may lead the model to respond to irrelevant background regions in the image. Therefore, background removal is expected to reduce the model’s sensitivity to irrelevant background regions. In this study, we examined the impact of background removal on the classification performance of transfer learning models applied to images of mounting hardware for lighting fixtures in highway tunnels. The results suggest that background removal may contribute to improving the classification performance of these models. for performance improvement in models with constrained feature extraction capability.

In recent years, machine learning has been increasingly applied to rental price estimation in the real estate domain. However, most existing studies are based on structured tabular data. In practice, floor plan images provide important visual cues often emphasized by prospective tenants. This study proposes a semantic segmentation approach to extract such features from floor plan images for rental price estimation. Specifically, we integrate terms into the loss function of a SegFormer-based model to predict room types that are semantically similar but mutually exclusive (e.g., bath, toilet, unit bath). These terms enforce class exclusivity during prediction and improve segmentation consistency across visually confusing room types.

In recent years, retrieval-augmented generation (RAG) using large language models (LLMs) has attracted significant attention. However, many real-world documents contain structured content such as tables and images, which present challenges for conventional RAG systems. This study proposes a method for generating unified structured documents by combining OCR, layout analysis, and LLMs to preserve structural information. Experimental evaluations using university AI guidelines demonstrate that our approach improves the accuracy of RAG-generated outputs.

Heat exchangers play a critical role in plants and industrial facilities, and effective flaw detection techniques are essential to maintain their integrity. This study aims to enhance flaw detection efficiency by analyzing waveform data obtained from eddy current testing using machine learning algorithms. Data preprocessing and feature extraction were performed, followed by a comparison of multiple algorithms, including Support Vector Machines, Random Forests, and Neural Networks. The results revealed variations in detection accuracy based on flaw types and locations, enabling the identification of highly accurate and reliable models. These findings are expected to improve the efficiency and reliability of non-destructive testing for heat exchangers.

Local railways are facing a decline in users, and additionally, they are challenged by the difficulty of accurately grasping actual usage conditions and latent demand—information that is essential for formulating policies aligned with local realities. This study proposes a method to address this issue by leveraging social media (SNS) data and large language models (LLMs) to conduct large-scale analyses of users’ raw voices related to local railways. Specifically, we develop a technique to automatically extract usage purposes from SNS posts and apply it to multiple railway lines with differing demand characteristics for comparative analysis. Through this analysis, we aim to clarify differences and commonalities in demand patterns across lines, and particularly to visualize latent demand and usage characteristics in low-demand local railways. This method is expected to contribute to the formulation of effective policies that are grounded in the actual conditions of local railway systems.

In recent years, with the advancement of LPWA (Low Power Wide Area) communication technologies and the increasing use of renewable energy, there has been growing interest in ICT-based solutions to address the escalating issue of wildlife damage. Among these, animal species identification technologies play a crucial role not only in improving the efficiency of trap-based capture methods but also in devising effective deterrent strategies such as repelling techniques. This study aims to develop a sustainable wildlife damage prevention framework for agricultural lands in mountainous and hilly regions, focusing on the development of core animal classification technology. Since many wild animals appear predominantly at night, identification is carried out using infrared (IR) images—rather than standard color images—captured by trail cameras and similar devices. Specifically, this research explores the application of deep learning techniques to classify animal species from IR images and evaluates the required amount of training data, classification accuracy, and the feasibility of practical deployment.

In defining fuzzy sets, we focus on the values of membership but pay little attention to the values of the universe of discourse (variable). For example, consider the definition of the fuzzy set "large number" on real numbers. When we consider 50, we expect the membership values for 49, 50 and 51 to be clearly different near 200. We, however, expect the membership values for 199, 200 and 201 to differ only slightly for values distant from 50. Since it is often the case that the membership values of distant values are usually defined as 1 or 0, we take little notice of them and are not so concerned about them. However, the relationship between the universe of discourse and memberships differs between the neighborhood of the considered value and its distant values, and they change smoothly from one to the other. In this paper, we consider several definitions of the membership function to avoid these differences.

This study aims to investigate approaches toward automating the analysis of handwritten waveform recordings of “rhythm lines” in the Kestenberg Movement Profile (KMP), a movement analysis method grounded in psychoanalytic theory. Tension Flow Attributes (TFA), one of the analysis categories of KMP, has six items, and in previous studies, a fuzzy set-based feature judgment support tool has been developed to judge four items. In this paper, we focus on the remaining two items, “Flow Adjustment (meandering of lines)” and “Even Flow (parallelism with Neutral Line),” and investigate a method for quantitatively extracting and judging the shape features of the rhythm lines. In the proposed method, an algorithm was introduced to objectively evaluate zigzagging and parallelism by calculating the amount of change, direction, parallelism, and other features from the waveform coordinate data using Microsoft Excel. The results showed that quantitative analysis support is possible and suggested that the proposed method contributes to the establishment of a foundation for objective analysis of all six TFA items.

As camera-equipped IoT devices proliferate in bandwidth- and power-constrained environments, the need for ultra-lightweight image compression has become more important than ever before. Recently, MCUCoder achieved efficient adaptive bitrate compression on microcontrollers using an encoder with only 10. 5K parameters. However, that model exhibits limited robustness against dynamic transmission control based on both image content and communication bandwidth. In this paper, we propose Fuzzy-MCUCoder, a model designed to achieve more effective transmission by considering factors such as bandwidth and image complexity. In the proposed model, a fuzzy inference system is incorporated into the encoder that decides the number of channels of the latent representation to be transmitted. As evaluated on the ImageNet dataset, Fuzzy-MCUCoder achieves an improvement in image reconstruction performance under the same transmission time, compared to prior work.

People with visual impairments often have difficulty perceiving their surroundings, which increases the risk during walking. In recent years, systems that provide voice guidance using sensors and image recognition have been developed. However, delivering intuitive, timely, and flexible instructions remains a challenge. This study proposes a walking support system based on fuzzy control, which integrates ambiguous and multivariate information to generate context-appropriate voice instructions. Fuzzy control is expected to enhance the clarity and naturalness of the instructions, as well as the user’s sense of safety.

This study proposes a novel fuzzy entropy-based evaluation framework for assessing emotional understanding capabilities of Large Language Models (LLMs) in Japanese literary texts. Unlike conventional discrete classification methods, our approach quantifies emotional ambiguity using fuzzy membership functions (Low, Medium, High) and measures evaluation uncertainty through fuzzy entropy ranging from 0.2 to 1.8. We conducted comprehensive experiments with 36 LLM variants across four distinct personas with systematically controlled temperature parameters (0.1-0.9), collecting 4,227 emotional evaluation data points with 97.8% completeness. The analysis revealed significant diversity in emotional understanding mechanisms across developer groups, with positive emotions showing up to 27.2-point differences between companies (Alibaba: 83.0±7.6 vs. xAI: 55.8±24.3). Language-numerical consistency analysis demonstrated substantial variations from 0.554 (Llama-4-Scout-17B) to 0.085 (Qwen3-235B-A22B-FP8), indicating a 6.5-fold difference in evaluation coherence. Our persona-temperature correlation analysis confirmed significant relationships (r = 0.73, p < 0.001), validating the theoretical foundation of cognitive diversity control in LLM evaluation systems.

In this study, we explored the possibility of modeling human decision-making processes using AI based on the framework of the Analytic Hierarchy Process (AHP). Specifically, we focused on the grading of essay assignments and analyzed the structure of evaluation criteria from the grading results using tools such as GPT-4o. As a result of applying three methods―thematic method, holistic method, and comparative method―for selecting evaluation criteria, it was revealed that the thematic method was the most effective. Furthermore, by introducing regression analysis into the construction of the evaluation criteria, we demonstrated that the grading process could be reproduced with high accuracy. In the future, we aim for the social implementation of AI-based decision support systems through further case analyses.

This study proposes a framework to build personalized support relationships in the form of pairs, focusing on individuals’ intrinsic skills rather than external attributes. It aims to identify members who require support and select suitable supporters to enhance overall productivity and collaboration. By introducing Interval DEA and Interval IDEA, the study enables multifaceted talent analysis and contributes to sustainable team management through tailored support strategies.

In recent years, in order to promote women's participation in STEM fields, there has been discussion and implementation of measures such as creating new quotas for women in university entrance exams, but it seems that few educators have actually listened to the voices of the female students who are the target of these measures. In this paper, we will consider the results of a survey conducted mainly among female students in science faculties at national and public universities and provide a discussion on measures that can lead to an increase in the proportion of female students in DE&I sessions.

At Panasonic Connect, we promote DEI（Diversity, Equity and Inclusion) as a key management strategy, focusing on respect for human rights and enhancing corporate competitiveness. Our CEO, along with other top executives, actively communicates this commitment both internally and externally. Additionally, DEI Champs -leaders responsible for DEI initiatives within each department- tailor their activities to the unique characteristics of their respective teams. This decentralized approach is a significant feature of our DEI promotion efforts. In this paper, we will introduce specific examples of our DEI initiatives, including relevant case studies that highlight our progress and impact.

The Japan Research Institute is a company with three functions: a think tank, consulting, and IT solutions, and it engages in various business activities. It considers human resources to be its most important asset. The consulting division has a system that allows involvement in both public and private sector projects, approaching contemporary Japanese issues from a multifaceted perspective. It emphasizes the diversity of consultants, promoting the development of expertise, skills, and freedom in working styles. As social issues become more complex, the demand for consulting is increasing, and it is considered important to strengthen relationships with external experts and respond to changes in the future.

Machine learning has been proposed for a wide range of applications, including materials development, medical diagnosis, and production and safety management, as it enables the extraction of useful patterns from complex data and facilitates prediction and optimization. To improve the performance of predictive and optimization tasks, enhancing the accuracy of machine learning models is essential. Enhancing model accuracy generally requires increasing the amount of training data. However, in many cases, the collected data cannot be used directly and must be converted into a format suitable for machine learning. In particular, image data often requires preprocessing such as cropping of target regions, which, when done manually, demands a significant amount of time and effort. In this presentation, we report on the implementation of automatic image cropping using YOLO-based object detection as an approach to address this issue.

Materials Informatics is a research and development field that applies information science technologies to materials development, aiming to discover new materials and improve the efficiency of materials design. In this presentation, we introduce an application of machine learning to materials informatics through a case study on the prediction of gallium crystal structures. We also discuss future challenges and prospects in this field.

We discuss parameter estimation of decaying oscillations using autoencoders. Decaying oscillations commonly occur in many physical systems and analyzing them can reveal the characteristics of the underlying physics. Rapidly decaying signals are challenging to analyze using conventional methods, and observational data often contain noise. Under such conditions, machine learning offers the potential for efficient and high-precision parameter estimation. In this study, we utilize the latent space of autoencoders to estimate the frequency, phase, and decay time of noisy decaying oscillations, and evaluate the estimation accuracy. Building on these findings, we aim to apply this approach to parameter estimation of ringdown gravitational waves in the future.

In the field of off-policy evaluation, methods have been proposed that use reward estimation models learned from data to predict rewards in unobserved domains. However, the data used for learning depends on the action selection probabilities of the policy used to obtain the data, and the model’s prediction accuracy may deteriorate due to selection bias. This is because variables that influence policy action selection also influence the results, leading to spurious correlations caused by confounding factors that are reflected in the prediction model. Therefore, this study aims to construct a reward estimation model based on causal relationships rather than correlation-based prediction models. As the first step, we constructed a causal graph from real data using the Peter-Clark algorithm, one of the causal exploration methods. Additionally, we analyzed the constructed causal graph and explored methods for applying it to reward estimation models.

In greenhouse cultivation, maximizing productivity requires assessing and understanding plant growth through various sensors and appropriately controlling the cultivation environment based on the measurement results of plant growth. In this study, we present an approach to measuring plant growth using first-person view images captured with a grower’s head-mounted camera and IMU sensor. The experimental results show that our approach has the potential to serve as a low-cost and easy-to-use method for assessing plant growth in greenhouse environments.

In welfare and employment support settings, there is a growing need to visualize the safety of activities and their rehabilitative effects.This study focuses on horticultural activities and proposes a method to detect contact with plants, aiming to support understanding of task content and provide instructional guidance.A two-view video setup was employed to develop and evaluate the proposed contact detection method and to explore viewpoint-based complementation.

In the food service industry, promoting repeat visits from existing customers is regarded as a cost-effective strategy, and accurately predicting such behavior is considered valuable for improving services and promotional efforts. However, repeat behavior depends on a wide range of factors, making it difficult to achieve high prediction accuracy using conventional heuristic approaches. Although machine learning techniques have been increasingly applied in recent years, investigations into models with higher generalization performance remain insufficient. Therefore, this study constructs multiple machine learning models to predict customers’ intention to revisit ramen restaurants in the Kanto region, using input features such as service experience and location conditions, and conducts a comparative evaluation of the models’ predictive performance. The findings are expected to contribute to the establishment of practical prediction methods for repeat customers in the food service sector.

In recent years, the issues faced by local communities have become increasingly complex, and the importance of accurately understanding residents' opinions and needs has been growing. In the analysis of resident questionnaires aimed at identifying the issues that residents consider important in their communities, cross tabulation, which aggregates combinations of two questions, is commonly used as a method to clarify inter-question relationships. However, cross tabulation has the limitation that it becomes difficult to interpret the results and discover important relationships as the number of question combinations increases. This study proposes a method that treats multiple cross tabulation results as tensor data, extracts major response patterns based on nonnegative CP decomposition, and clusters these patterns. We also report the analysis results obtained by applying the proposed method to actual resident questionnaire data.

Achieving high-precision pose estimation remains dependent on the labor-intensive annotation of keypoints, which poses a significant challenge in terms of human resource cost. This study aims to mitigate such costs by investigating the applicability of knowledge acquired in a source domain to a previously unseen target domain. Leveraging model editing techniques based on Task Vectors, we transfer the knowledge obtained from human pose estimation to the task of animal keypoint estimation. The edited models are evaluated in terms of both prediction accuracy and tendency, thereby assessing the effectiveness of domain adaptation across different target categories.

In recent years, the use of data in sports has accelerated. For instance, in the domain of soccer, there is ongoing research aimed at leveraging tracking and event data to facilitate tactical analysis and performance evaluation. However, the process of acquiring such data frequently necessitates the laborious task of manual annotation or the utilization of costly, specialized equipment. In particular, event data is generally collected by human operators who watch match videos and annotate events using proprietary software, which involve considerable time and cost. In this research, we propose a method to automatically detect passes from broadcast soccer videos and efficiently obtain event data. In proposed method, extracts the positional information of players and the ball from the video, as well as the image features of the player closest to the ball, and uses this information to estimate whether a pass is occurring in each frame. By focusing on broadcast videos, the potential exists to undertake large-scale analyses using past and globally distributed match videos.

In recent years, artificial intelligence (AI) has found applications in various fields, including sports. Technology that uses machine learning to analyze camera images and provide real-time information has attracted particular attention. In basketball, table officials (TO) work with referees to assist with the game. They are responsible for tasks such as recording the score, managing fouls, and operating the timer. Currently, all of these tasks are performed manually, requiring at least four TO for an official game. Interruptions sometimes occur due to human error. In this research, we developed an AI-based system that recognizes foul gestures of basketball referees to assist TO with their duties. Using videos in which the referee was clearly visible, we extracted skeletal information with MMPose and classified foul gestures based on time-series data. The system achieved 95% accuracy in classifying the type of foul and 74.4% accuracy in recognizing the jersey numbers of the players who committed the foul. This research is expected to support TO tasks and ensure fair game management. However, further improvements, especially in jersey number recognition accuracy, are necessary for practical implementation.

This study proposes a method that leverages the geometric relationship of acceleration vectors for gait recognition. We use a public dataset in that acceleration sensors were attached to the right pelvis and left thigh of walkers. A low-pass filter based on 95% power distribution was applied to each body part to extract the main motion components. As features, squared magnitudes of the acceleration vectors and dot and cross products of two simultaneously measured acceleration vectors were calculated to represent the characteristics of individual walker. Using these features, a classification experiment was conducted with a Random Forest (RF) algorithm.

In the RoboCup Middle Size League (MSL), self-localization based on white line information is crucial for the strategic plays of autonomous mobile robots. Conventionally, parameters for white line extraction were adjusted manually, leading to inconsistent accuracy depending on operator proficiency. To address this, we developed a system using deep learning to automatically recognize white lines with high precision from the robot’s onboard omnidirectional camera images. This method eliminates operator-dependency in parameter tuning and aims to achieve stable, highly accurate self-localization.

It has become difficult for users to find music that matches their preferences with the increasing number of songs due to the development of subscription services in recent years. To address this problem, in this study, we extract music features and use them to generate impression images and develop a recommendation system. As a result, we developed a recommendation system using the music features. We used ‘Stable Diffusion’ to generate images that represent the mood of the music. Additionally, we created ‘Animated GIF’ as another method of dynamic.

This study aims to improve the dialogue continuity problem in non-task oriented dialogue systems by using humor. The proposed method considers logical incompatibility, structural incompatibility, and resolution in the “incompatibility-resolution model” and generates short stories using mishearing. In an evaluation experiment using a prototype system, the proposed method was able to generate unexpected speech for the user by generating humor, but issues were found regarding the “timing of generating short stories”, "funniness” and “dialogue continuity”.

This study develops a system to predict, with high precision and in real-time, the occurrence of difficulty in character identification during web browsing, based on gaze data. Specifically, the system leverages fixation duration, which evolves incrementally, and employs a reinforcement learning algorithm based on SARSA, to evaluate the occurrence of the difficulty at each step. Since fixation durations caused by character identification difficulty are not necessarily longer than those resulting from other factors, establishing a reliable threshold for character magnification is difficult. Nevertheless, the system must refrain from magnifying characters when users do not feel them difficult to identify. Therefore, this study introduces saccadic velocity and amplitude as two external parameters, categorizes them into distinct groups, and calculates the Q-value for each category pair, thereby enabling a precise determination of magnification thresholds.

Visually impaired people may encounter accidents such as colliding with obstacles or falling due to steps when going out. In particular, falling from a station platform poses a significant danger. One cause of such falls is losing their sense of position and direction. To address this issue and support visually impaired people in navigating safely, this study aims to develop a smartphone application that determines the vanishing point from camera images and tell that direction. Furthermore, this study will verify the usefulness of direction guidance by vanishing point in various environments and investigate the optimal detection methods.

In this paper, we consider a method to improve the multi-class logistic regression analysis. For the proposed method, we consider the characteristics of data matrix from the perspective of numerical analysis. We also consider the characteristics of data matrix from the perspective of eigenvalue distribution.

In semiconductor manufacturing, wafer maps are widely used to visualize and analyze the spatial distribution of defective chips for yield improvement. As modern production processes achieve high yields, defect patterns tend to become sparse and localized, making it increasingly difficult to detect subtle and unknown patterns using conventional methods. In this study, we propose a Graph-Based Clustering method for detecting small, localized defect patterns that are not classified into known categories. The proposed method calculates similarity between defect regions based on the structure and spatial features of chip clusters, and forms overlapping clusters to capture potential anomalies. To evaluate the robustness of our approach with respect to defect location, we created 13 different datasets by dividing the wafer into 13 regions and injecting artificial defects with the same characteristics into each region. Experimental results show that the proposed method maintains high F1 scores across all regions, demonstrating robustness to the location of defects. These findings suggest that our method is effective in detecting subtle and unknown defect patterns in high-yield wafer maps, regardless of their position.

Soldering tasks are still often performed manually, and variations in product quality due to physical characteristics and skill levels of operators remain a challenge. In this study, we applied K-means unsupervised clustering to the right-hand second joint angle data computed using MediaPipe to extract patterns in grip styles. By comparing cluster distributions and analyzing cosine similarity, we evaluated differences based on physical characteristics. The results suggest that soldering grip styles may differ depending on physical traits, with the dominant eye showing particularly significant influence. In the future, we aim to utilize the obtained grip vectors to support beginner training and skill acquisition.

In this study, we developed a system that automatically acquires and shares environmental data collected from agricultural sensors by utilizing Robotic Process Automation (RPA) and Google Apps Script (GAS) in a web application. The sensors send data via the SORACOM platform, and traditionally these data were used solely by researchers. Our system employs RPA to periodically download the latest data from SORACOM and automatically updates and publishes it on a web application developed with GAS, allowing cooperating farmers and stakeholders to access real-time information. Furthermore, this application functions as a data infrastructure for artificial intelligence (AI)-based environmental data analysis, contributing to the preparation of an environment for future AI model development and application. The system is low-cost and highly versatile, with potential applicability to other fields.

In this study, we propose a control system that supports partial execution so that beginners can efficiently perform trial-and-error programming on a small hydraulic excavator using Scratch. In conventional systems, it is necessary to redo all instructions from the initial posture for each trial, and tests to correct parts of the program take a long time and are inefficient. This system introduces two execution modes: initial posture-based and current posture-based, and allows users to specify the execution range in a block sequence using a dedicated block. In the initial posture-based mode, the system can quickly and automatically restore the required machine posture by reverse execution or re-execution after initialization. We evaluated the effectiveness of the system through a case study comparing different execution modes, and confirmed that partial execution improves execution time while ensuring reproducibility. The current system does not support simultaneous operation of multiple actuators, so we will continue to develop it in the future.