研究者詳細 - 吉田　香

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=105018923063&origin=inward

Enhancing Decision-Making System with Implicit Attitude and Preferences: A Comprehensive Review in Computational Social Science 査読有り国際誌

SILPANI Dian Christy, YOSHIDA Kaori

International Journal of Affective Engineering ( 日本感性工学会 ) advpub ( 0 ) 2024年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

Implicit attitudes and preferences can be dynamic and context-dependent, evolving over time and in different situations. In this study, we point out 4 main topics: Implicit attitudes, preferences, decision-making system, and computational social science, to introduce a new approach to developing a decision-making system. As a comprehensive review, we set the 3 research questions as the main considerations for this study. We collected 903,451 articles using 3 major database sources (ScienceDirect, ACM Digital Library, and IEEE Xplore), screened 5 times with different viewpoints, and surveyed 14 articles to answer the research questions. As a result of the comprehensive review, we conclude that it shows the possibility of enhancing the decision-making system with implicit attitudes and preferences in a computational social science manner.

DOI： 10.5057/ijae.ijae-d-24-00011

The Impact of Large Sample Datasets on Hand Gesture Recognition by Hand Landmark Classification 査読有り国際誌

AHMAD Khawaritzmi Abdallah, SILPANI Dian Christy, YOSHIDA Kaori

International Journal of Affective Engineering ( 日本感性工学会 ) 22 ( 3 ) 253 - 259 2023年07月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

Today, almost all aspects of life are influenced by technology to simplify these aspects of life. Researching how humans and computers interact is very important to build a good model for better human-computer interaction in the future. Using technology such as computer vision, we can now collect the information we are looking for directly from a human. Humans can use many kinds of modalities to interact with computers. Hands are perhaps the largest source of body language information after the face. To understand the gesture’s meaning, we can use MediaPipe Hands, developed at Google LLC, as a method to track and recognize human hands. However, if we want to understand some kinds of hand gestures using MediaPipe Hands, we need to create a condition using if-else manners. This research tried to collect the wide variety of each hand gesture using the 21 key points in x, y, and z coordinates as a feature. We chose Support Vector Machine (SVM) and Artificial Neural Network (ANN) as classifiers to validate the impact of large sample datasets. This research found that ANN is the best among all the methods we used as a classifier method for the 3D value of 21 key points from the hand skeleton. The accuracy and F1-score from ANN were 98.4% accuracy and 98.2% F1-Score, respectively, representing the best performance for each class of all the methods we used in this research.

DOI： 10.5057/ijae.ijae-d-22-00022

その他リンク： https://www.jstage.jst.go.jp/article/ijae/22/3/22_IJAE-D-22-00022/_pdf

特集「伝統芸能・伝統工芸と人工知能」にあたって

佐久間洋司, 森友亮, 吉田香

人工知能 ( 一般社団法人人工知能学会 ) 41 ( 4 ) 412 - 413 2026年07月

Vision-language models meet hand gesture recognition: towards multimodal and intelligent human-computer interaction 査読有り国際誌

Indasyah E., Yoshida K.

Visual Computer 42 ( 9 ) 2026年07月

担当区分：最終著者記述言語：英語掲載種別：記事・総説・解説・論説等（学術雑誌）

Hand Gesture Recognition (HGR) has become an important means of enhancing natural interaction with computers. However, traditional vision-only systems are limited by closed-set classification and lack firm semantic grounding. The recently introduced large vision-language models (VLMs), such as CLIP, BLIP, Flamingo, and GPT-4V, provide open-vocabulary recognition, cross-modal retrieval, and gesture captioning by learning common visual-textual representations from large image-text corpora. This survey, therefore, synthesizes approaches that fall within the general scope of attempts to inject vision-language models (VLMs) into the human-robot repertoire and neatly fits them into a three-dimensional taxonomy consisting of: learning paradigms which range from zero-shot, few-shot, prompt-based, retrieval-augmented to parameter-efficient fine-tuning; task objectives spanning classification, captioning, retrieval, and multimodal reasoning; input modalities including static images, video, RGB-D, and textual prompts. We analyze how VLMs enable the convergence of linguistic and perceptual representations toward enhanced gesture understanding, adaptability, and explainability. Building on this taxonomy, we describe the limitations of current datasets and benchmarking practices for gesture-language alignment, which manifest most significantly in cases of small datasets across linguistic diversity and varying cultural practices. We next map the open problems and research gaps toward building multimodal HGR systems that enable semantically grounded, culturally informed, and context-sensitive intelligent interactions.

DOI： 10.1007/s00371-026-04518-7

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=105043785147&origin=inward

Study of Emotional Word-Color Association in Tea-Tasting Experiment 査読有り国際誌

Rukka F.M., Silpani D.C., Yoshida K.

Lecture Notes in Networks and Systems 1904 LNNS 227 - 241 2026年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Colors are strongly linked to emotions, yet different types of emotional words, such as labels (e.g., “happy”) and laden words (e.g., “birthday”) might have different color associations. This paper will discuss and analyze how people associate colors with different word types to identify patterns. A tea-tasting experiment was conducted, where participants tasted different types of tea, sharing their impressions, choosing words, and describing colors. The colors were analyzed using the HSV (Hue, Saturation, Value) model to capture the intensity and consistency of word–color associations. A Chi-square test shows a significant relationship between word valence and color selection (χ2=148.33,df=40,p<0.0001), meaning that specific types of emotional words are consistently linked with particular color ranges. Label-positive words showed strong clustering on the color wheel. In contrast, laden-positive words showed broader, more varied distributions. Neutral words had the widest spread. These results highlight meaningful patterns in how people map emotions to colors and suggest applications in affective computing, sensory research, and consumer product design. While this study offers insights, limitations such as a small sample size and cultural or individual biases suggest the need for broader, more flexible future research, such as using a color wheel directly to explore more emotional responses.

DOI： 10.1007/978-3-032-20897-2_16

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=105045689406&origin=inward

Class Balancing in Speech Emotion Recognition: A Comparative Study of Data and Feature Level Approaches 査読有り国際誌

Muhammad Amil Siddik Andi, Yoshida Kaori, Lago Paula

International Journal of Activity and Behavior Computing ( 九州工業大学ケアXDXセンター ) 2026 ( 1 ) 1 - 15 2026年01月

記述言語：英語掲載種別：研究論文（学術雑誌）

Class imbalance often causes Speech Emotion Recognition (SER) systems to perform well on majority emotions while degrading recognition of minority classes. Although data-level augmentation and feature-level oversampling have been widely studied, their effectiveness across varying degrees of imbalance has not been systematically examined in a controlled setting. This study shows how imbalance severity influences the performance of class balancing strategies in SER. An imbalance-aware experimental framework is adopted in which imbalance severity is explicitly controlled within a single dataset. Using RAVDESS, three imbalance settings (IR≈1, original, and IR≈6, representing low to severe class imbalance) are constructed to isolate the effect of imbalance level, while EmoDB is used as an external validation dataset exhibiting natural class imbalance. Data-level augmentation, feature-level oversampling using SMOTE, and a hybrid strategy combining both are evaluated under a unified 1D-CNN pipeline with multi-seed experiments. The results show that SER performance consistently degrades as imbalance severity increases, with baseline macro-F1 decreasing from 0.92 (original distribution) to 0.88 (IR≈6). SMOTE shows unstable behavior under severe imbalance, as seen by increased performance variance. Data-level augmentation improves robustness but does not consistently dominate across imbalance settings. In contrast, the hybrid strategy achieves the most stable and effective performance, improving macro-F1 to 0.91 (IR≈6) and to 0.82 on EmoDB. The findings indicate that the severity of imbalance is crucial in assessing the efficacy of balancing measures and must be explicitly considered in the design and evaluation of SER systems.

DOI： 10.60401/ijabc.129

特集「地域産業を支える知の拠点」に寄せて

吉田香

感性工学 ( 日本感性工学会 ) 23 ( 3 ) 93 - 93 2025年09月

Knowledge-Infused Retrieval Boosts Few-Shot Hand Gesture Recognition on HaGRID with Vision-Language Model 査読有り国際誌

Indasyah E., Yoshida K.

Frontiers in Artificial Intelligence and Applications 411 396 - 413 2025年09月

担当区分：最終著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

Hand gesture recognition remains challenging, primarily due to its reliance on large-scale annotated datasets and the limited adaptability of existing models when encountering novel gesture classes. In this work, we propose to apply Adaptive Vision-Language Model (Adaptive-VLM). This lightweight, training-free framework utilizes only one image per class to recognize gestures on the HaGRID benchmark. Built upon the CLIP backbone, our approach incorporates symbolic knowledge-infused prompts, multi-prompt contextualization, and semantic exemplar ranking to improve few-shot generalization. Adaptive-VLM achieves a macro F1-score of 65.75% on the HaGRID test set (540 images) without any parameter fine-tuning, using 18 example images. It significantly outperforms the Random-VLM baseline (59.95%) and a ResNet-18 model fine-tuned for 10 epochs (4.09%) under the same data constraints. These findings highlight the effectiveness of combining structured domain knowledge and guided exemplar selection to overcome data scarcity in low-resource gesture recognition. Adaptive-VLM offers a promising direction for building adaptive and efficient HGR systems, especially in real-world human-computer interaction scenarios requiring rapid deployment with minimal data.

DOI： 10.3233/FAIA250540

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=105021110524&origin=inward

特集「AIとファジィの接点」にあたって

吉田香

人工知能 ( 一般社団法人人工知能学会 ) 40 ( 3 ) 266 - 268 2025年05月

矢鳴虎夫先生（BMFSA 第6 代会長（1999-2001）・名誉会員）を偲ぶ

和辻直, 白濵成希, 宮本和典, 吉田香, 有田清三郎, 山川烈, 内野英治, 渡邉志

バイオメディカル・ファジィ・システム学会誌 ( バイオメディカル・ファジィ・システム学会 ) 26 ( 2 ) 59 - 66 2025年02月

BLE Signal Reception and Localization Performance with Varying Receiver and Beacon Setups 査読有り国際誌

Benaissa B., Hendrichovsky F., As M., Yoshida K.

Future Internet 17 ( 2 ) 2025年02月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

This paper examines the performance of Bluetooth Low Energy signal reception for indoor localization by analyzing the interactions between gateways, beacons, and receiver placements. The study investigates the effect of different BLE beacon placements on signal strength and localization accuracy. It evaluates ten receiver ceiling-mounted and wall-mounted configurations, as well as five beacon body positions: shoulder, front pocket, back pocket, and wrist. A dataset comprising 2700 data points was collected and localization accuracy was assessed using a Radial Basis Function-based methodology. The results demonstrate that ceiling-mounted gateways offer more stable signal strength and enhanced localization accuracy compared to wall-mounted gateways. The findings highlight the significance of optimizing both gateway positioning and body placement to improve the performance of BLE-based indoor positioning systems.

DOI： 10.3390/fi17020054

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85219182448&origin=inward

私のブックマーク：感性情報処理

吉田香

人工知能 ( 一般社団法人人工知能学会 ) 40 ( 1 ) 1080 - 1087 2025年01月

Dynamic Hand Gesture Recognition by Hand Landmark Classification Using Long Short-Term Memory 査読有り国際誌

Ahmad K.A., Higashi T., Yoshida K.

Pertanika Journal of Science and Technology 33 ( S2 ) 73 - 84 2025年01月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

Hand gestures are a valuable modality for human-computer interaction, conveying information that can be used as input. Dynamic hand gestures, prevalent in real-world scenarios, necessitate considering temporal factors such as gesture initiation, termination, and frame sequence. A Long Short-Term Memory (LSTM) based recognition model was proposed to address this challenge. Data availability for dynamic hand gesture research is a significant hurdle. The dataset introduced by Fronteddu et al. provides 27 classes of dynamic hand gestures, serving as a suitable training resource. MediaPipe Hands, a computer vision framework, was leveraged to extract keypoints from each frame, capturing spatial features fed into the LSTM model. Experiments were conducted to determine the optimal dropout rate for the LSTM model. Results indicated that a dropout rate of 70% yielded the highest accuracy, achieving up to 98.53% validation accuracy and 99.71% test accuracy. These findings demonstrate the effectiveness of the proposed LSTM-based recognition model for dynamic hand gestures. Future research could explore integrating other deep learning techniques, such as attention mechanisms, to enhance the accuracy and robustness of dynamic hand gesture recognition systems. Additionally, investigating the application of the proposed model in real-world scenarios, such as virtual and augmented reality, would be valuable in assessing its practical utility.

DOI： 10.47836/pjst.33.S2.05

その他リンク： https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85219544263&origin=inward

Fuzzy Sentiment Analysis for Hand Gesture Interpretation in Words Expression 査読有り国際誌

Rukka Faizah Mappanyompa, Silpani Dian Chisty, Yoshida Kaori

Proceedings of Computer Information Systems, Biometrics and Kansei Engineering ( 生体認証と感性工学に関する国際会議 ) 2025 ( 0 ) 106 - 119 2025年01月

担当区分：最終著者,　責任著者記述言語：英語掲載種別：研究論文（国際会議プロシーディングス）

This paper will discuss the interpretation of hand gestures using words with the fuzzy
sentiment analysis approach in the tea-tasting experiment. In the experiment, participants were
asked to drink tea, give a hand gesture, and choose three words that could interpret the hand
gesture. The results show that hand gestures cannot be categorized as fixed sentiment labels, considering
the meaning can differ depending on each preference. Generally, traditional sentiment
categorizes labels as positive, negative, or neutral, which fails to capture human preferences.
Therefore, we propose that the fuzzy sentiment model be used to solve the traditional problem,
which provides a more flexible approach. Our findings show that interpreting hand gestures such
as positive is not always positive but can be neutral-positive and vice versa. The results show that
hand gestures alone are insufficient to determine sentiment; word expression is required to capture
sentiment's different and complex aspects.

DOI： 10.69333/icbake.2025.0_106

Static Hand Gesture Recognition Based on Multi-Feature Extraction with Context for User Intension 査読有り

Ahmad K.A., Silpani D.C., Yoshida K.

ICIC Express Letters, Part B: Applications ( ICIC International 学会 ) 15 ( 8 ) 779 - 786 2024年08月

担当区分：最終著者記述言語：英語掲載種別：研究論文（学術雑誌）

Hand gesture is a source of information that contains much information, which has the advantage that this modality does not contain private information as much as information obtained from facial modalities. However, obtaining information from hand gestures still encounters several challenges, one of which is the diversity of gestures within one target class which is influenced by many factors such as the culture and personality of the individual. In this study, we propose a method of combining hand gesture features in the form of keypoints features, convolutional neural network features, and manually generating captions as contexts to improve model performance so that the model can classify hand gestures based on the intent of the hand gesture with context support. As a result, the CNN model, which initially experienced over-fitting when trying to classify hand gestures based on the intent of the hand gesture, can be overcome by this method. Running results show a significant increase in accuracy over training models that just use uni-features, with a model having a training accuracy of 99.95% and a validation accuracy of 99.74%. The final result of the model obtained the ability to understand 6 classes of hand gestures made in 3 classes based on the intent of the hand gestures exhibited when participants try the taste of tea, namely Like, Dislike, and Neutral (So-so).

DOI： 10.24507/icicelb.15.08.779