Faculty Profiles - TOKUNAGA Terumasa

Conference Prsentations (Oral, Poster) -

Division display21 - 40 of about 79／ All the affair displays >>

外部視覚注視機構による深層学習の異常検出能力と注視領域生成方法の関係について

城間亮, 早川季寿, 中西慶一, 徳永旭将

第26回画像の認識・理解シンポジウム(MIRU2023) 2023.07 電子情報通信学会PRMU研究会

パッチ画像を用いた半教師あり学習に基づくセグメンテーションフレームワーク

中⻄慶一, 徳永旭将

第26回画像の認識・理解シンポジウム (MIRU2023) 2023.07 電子情報通信学会PRMU研究会

Heavy Snow Cloud Detection in Satellite Images Based on Semi-Supervised Image Segmentation

Lin Magari, Terumasa Tokunaga, Kazue Suzuki

2023.05

Prediction of off-angle from GaN surface morphology images using deep learning

The 9th Meeting on Advanced Power Semiconductors Division 2022.12

A rule-based anomaly detection method using binary segmentation as preprocessing

Kazumichi Tanaka, Sansei Hori, Keiichi Nakanishi, Terumasa Tokunaga

The 4th International Symposium on Neuromorphic AI Hardware 2022.12 Research Center for Neuromorphic AI Hardware

Event date： 2022.12.13 - 2022.12.14 Language：English Country：Japan

Semiconductor visual inspection systems are required to further improve productivity and quality control in proportion to the increasing demand for semiconductor products. In machine vision systems, the combination of conventional rule-based anomaly detection methods, which are characterized by their high explanatory power for the basis of judgment, and deep learning methods, which are characterized by their high extraction and classification accuracy for features such as complex shapes, is an effective means from the perspective of productivity and quality assurance.
We propose a rule-based anomaly detection method that uses binary segmentation [1] as a preprocessing step. Figure 1 shows the inspection flow model of the proposed method. Binary segmentation is applied to the input image for regions that are difficult to distinguish from defect regions by clear human definition and that are a cause of over-detection with characteristics such as shape and pixel value. The extracted regions are then combined with the input image as a mask. Then, a rule-based abnormality detection inspection is applied to the synthesized image.

GL-CANomaly: Global and Local adversarial image completion networks for ANomaly detection

Takara Ishimoto, Keiichi Nakanishi, Sansei Hori, Terumasa Tokunaga

The 4th International Symposium on Neuromorphic AI Hardware 2022.12 Research Center for Neuromorphic AI Hardware

Resorvoir-based neuron model to emulate cellular neural activity responding to odor in C. elegans

Ryosuke Ishibashi, Takumi Nakamura, Noriko Sato, Takeshi Ishihara, Yuichiro Tanaka, Hakaru Tamukoh, Terumasa Tokunaga

The 4th International Symposium on Neuromorphic AI Hardware 2022.12 Research Center for Neuromorphic AI Hardware

Event date： 2022.12.13 - 2022.12.14 Language：English Country：Japan

The nervous system of Caenorhabditis elegans (C. elegans) realizes sensory-motor transformation for many odorants despite it consists of only 302 neurons. Recently, we have developed a novel live-cell imaging technique based on genetically encoded voltage and calcium indicators. Our imaging system enables simultaneous measurements of voltage and Ca2+ responses to specific sequential odor stimulation at a single cellular level. The series of experiments interestingly suggest that an individual neuron can sensitively and quantitatively express time-varying environmental information quite unlike the formal neuron. In this study, we attempted to model a single-neuron of C. elegans in a data-driven manner. As a pilot study, we validated whether voltage responses in AWA, which is one of the olfactory sensory neurons of C. elegans, can be reproduced by a simple echo state network (ESN). For input data, we used synthetic square waves that imitate sequential odor stimulation in in-vivo experiments. For target signals, actual measurements of voltage responses in AWA were used. The prediction test by the trained ESN showed that the transient variations of membrane potentials can be successfully reproduced while we failed to reproduce quasi-stationary depolarizations during odorant stimulation (Figure 1). The result warrants further research on more realistic ESN models including multiple voltage dependent ion channels.

Distribution-Free Semi-Supervised Cost Function with a Class-Prior Probability

Keiichi Nakanishi, Terumasa Tokunaga

The 4th International Symposium on Neuromorphic AI Hardware 2022.12 Research Center for Neuromorphic AI Hardware

Event date： 2022.12.13 - 2022.12.14 Language：English Country：Japan

Supervised learning requires a large amount of labeled data to obtain high generalization performance. The labeled data is created by annotation. However, the annotation has a lot of problems, such as high cost. In this study, we propose a novel semi-supervised cost function for using unlabeled data effectively, which consists of two important terms. There is a class-prior probability penalty and an extended Focal loss function. A class-prior probability prevents overfitting to labeled data by utilizing prior knowledge for a percentage of positive samples over unlabeled data. We performed binary image segmentation to evaluate the proposed cost function's effectiveness. As a result, the proposed cost function performs flexible segmentation of objects with ambiguous shapes, because the output values of the boundary regions transition smoothly compared with Binary Crossentropy (BCE) and PNU Loss.

GL-CANomaly: Global and Local adversarial image Completion networks for ANomaly detection

2022.11

クラス事前確率を用いた分布を仮定しない半教師ありコスト関数とその応用

中西慶一, 片渕凌也, 堀三晟, 徳永旭将

第25回情報論的学習理論ワークショップ (IBIS2022) 2022.11 電子情報通信学会情報論的学習理論と機械学習研究会

半教師あり2クラス分類に基づく尤度を利用した粒子フィルタによる線虫の神経細胞追跡手法の提案

坂田大地, 中西慶一, 佐藤則子, 石原健, 徳永旭将

第11回生命医薬情報学連合大会 2022.09 日本バイオインフォマティクス学会, 日本オミックス医学会

膜電位イメージングデータの画像解析パイプラインの開発

Terumasa Tokunaga, Noriko Sato, Takeshi Ishihara, Keiichi Nakanishi, Daichi Sakata, takumi Namamura

線虫研究の未来を創る会 2022.08

A Class-prior probability regularization with an extended Focal Loss for efficient Semi-supervised classification

The 3rd International Sympojium on Neuromorphic AI Hardware

Atmospheric Riverによる南極域へのエアロゾル輸送(1)

鈴木香寿恵, 原圭一郎, 徳永旭将, 後藤大輔, 平沢尚彦, 山内恭

気象学会2021年度秋季大会日本気象学会

Event date： 2021.12.02 - 2021.12.08 Language：Japanese

近年，水蒸気輸送と豪雨(雪)には対流性の雲が連なって形成されるAtmospheric River(AR)が関連していると考えられるようになり，南極域においても極方向の水蒸気フラックス強化となる背景場とARの観測事例が報告されている[1]．そこで，昭和基地の降雪時にARと判別できる雲画像を用いたCNNによる自動識別に取り組んできた[2]．また，客観解析データを用いた全球規模のARとエアロゾル輸送の関連について報告がされており，Aerosol Atmospheric River (AAR)となって高濃度エアロゾルが輸送されることが示されている．本研究では，これまで行ってきたARによる水蒸気輸送だけではなく，陸起源と考えられる大気中微量物質の輸送も同時に捉え，大気による物質輸送過程を機械学習による予測モデルを構築することを目指す．まずは，ARとエアロゾル輸送の関連について2009年のブリザードイベントについて調べた．

LEA-Net: Layer-wise External Attention Network for Efficient Color Anomaly Detection

The 24th Information-Based Induction Sciences Workshop

CNNを用いた顔認識に対するマスク着用の影響と改善方法の検討

溝田十悟, 徳永旭将

第24回情報論的学習理論ワークショップ (IBIS2021) 電子情報通信学会情報論的学習理論と機械学習研究会

複数の脈波抽出領域と独立成分分析を用いたサーマルカメラによる非接触バイタルセンシング手法の提案

野見山陸, 徳永旭将

第24回情報論的学習理論ワークショップ (IBIS2021) 電子情報通信学会情報論的学習理論と機械学習研究会

言語の違いに頑健なText-to-Imageモデルの構築に向けた展望

仲地早司, 徳永旭将

第24回情報論的学習理論ワークショップ (IBIS2021) 電子情報通信学会情報論的学習理論と機械学習研究会

Aurora Image Segmentation with Deep PNU Learning

The 150 th SGEPSS General Assembly

Observation and simulation of C. elegans whole-brain neural activities

Event date： 2021.09.21 - 2021.09.23 Language：English

C. elegans is a model organism in which the structure (connectome) of the whole nervous system composed of 302 neurons has been determined. We performed whole-brain imaging by spinning disk confocal microscope combined with piezo objective positioner to obtain calcium imaging data of the whole head neurons. Further, we modeled the dynamics of neuronal ensembles based on the observed activity data and connectome data. As a result, we could perform virtual ablation of neurons or particular connections to gain insights into the information flow through the neural circuits.

　 PREV - NEXT 　

<TOKUNAGA Terumasa>

Conference Prsentations (Oral, Poster) -