ゼミナール発表

日時: 1月16日(月)3限(13:30-15:00)


会場: L1

司会: 河合 紀彦
久保 大輝 1651046: M, 1回目発表 自然言語処理学 松本 裕治
title: Annotation of Compound Functional Expressions in Japanese Universal Dependencies
Abstract: Universal Dependencies (UD) adopts dependency structure on a word basis, and in principle expresses a syntactic structure by a dependency relation which contains content words at least in one. Therefore, unless proper handling of functional expressions composed of a plurality of words is performed, a meaningful syntactic structure can not be obtained. In UD, the "mwe" corresponding to functional expressions is defined as a label of a dependency structure, but the current Japanese corpus (UD 1.2) does not annotate exhaustively. In this research, we aim to organize the framework of language resource development and syntactic analysis for syntactic analysis considering MWE by annotating compound words into UD even in functional expression.
language of the presentation: Japanese
 
太田 瑶子 1651023: M, 1回目発表 自然言語処理学 松本 裕治
title: Construction of haiku generation system using sequence-to-sequence model
abstract: There are many Japanese people who know the rules of haiku. However, sometimes It is not easy for them to make haiku that express what they want to say. In this research, I aim to construct a haiku generation system that allows people who are trying to make haiku to express what they want to say in haiku. In this presentation, I introduce the preceding research and report on future research plan.
language of the presentation : Japanese
発表題目: sequence-to-sequenceモデルを用いた俳句生成システムの構築
発表概要: 俳句のルールを知っている日本人は多い。しかし、いざ俳句を作ろうとしても言葉が浮かばず、思うように俳句を作ることができないことがある。本研究では、俳句を作ろうとしている人が、言いたい事を俳句で表現できるような俳句生成器の構築を目的とする。本発表では先行研究の紹介をし、今後の研究計画について報告する。
 
北川 智也 1651040: M, 1回目発表 情報基盤システム学 藤川 和利
title: (Paper Introduction) Improving the detection accuracy of unknown malware by partitioning the executables in groups
abstract: While the threats of cyber attacks increases, the importance of Anti-Malware study is increasing day by day. In this presentation, I introduce the study to classify unknown malware by two methods using Naive Bayes classifier, regular method and partitioned method. In the regular method, authors select the features by taking all datasets in one group and in the partitioned method, they select the features by partitioning the dataset in the range of file 5KB size. the study reports that the partitioned method detect the malware with ~8.7% more accurate than the regular method.
language of the presentation: Japanese
 
森口 和也 1651113: M, 1回目発表 生体医用画像 佐藤 嘉伸

title: Shape analysis and segmentation for abdominal organ with diseases

abstract:Abdominal organ segmentation from medical images is useful for medical diagnosis support.

Some organs is caused to transform by disease,and it makes difficult to segment them.

The purpose of this study is improving the extraction precision of segmentation.

Here,we report contents of relevant studies and transition of liver shape in hepatic fibrosis.

language of the presentation: Japanese

 
佐野 徳美 1651055: M, 1回目発表 生体医用画像 佐藤 嘉伸
title: Four-dimensional ultrasound imaging to analyze muscle function
abstract: Analysis of three-dimensional structure of muscles and its movement is important for sports science and rehabilitation. Ultrasound is widely used for analyzing muscle function, however field-of-view (FOV) of a common ultrasound probe is not large enough to analyze the entire region of an individual muscle, for example, in the case of forearm. In several studies, the FOV has been extended by mosaicking a series of 2D images or using an optically or magnetically tracked probe. However, freehand sweeping does not ensure the accuracy of data collection, so an image quality is not stable. In this study, we developed a mechanical guide system for reconstructing large 3D ultrasound volume with high accuracy. In the experiments, we compared the reconstruction accuracy of the freehand system and the proposed linear guide system using a block shape phantom. Furthermore, we reconstructed forearm of a volunteer and confirmed three-dimensional of the muscle shape and muscle fiber while closing the hand.
language of the presentation: Japanese
発表題目: 筋機能解析のための4次元超音波画像再構成
発表概要: スポーツ科学やリハビリテーションの分野において,筋肉内部の構造や動態を解析することは重要であり,そのためには筋肉が見えやすく動態の撮像に適した超音波診断装置を用いるのが一般的である.一方,一般的な超音波プローブは撮像範囲が狭く筋肉全体を撮像できないため,二次元画像処理による撮像範囲拡張や三次元位置計測を用いた広範囲ボリュームの再構成が行われている.しかしながら,再現性の乏しいフリーハンドの掃引では常に一定した画質を得ることは難しく,動態解析には適さない.そこで,我々は筋肉の三次元動態の解析を目的とし,広範囲かつ安定した撮像が可能な機械式リニアガイドシステムを構築した.本発表では,解析の前段階としてまず構築したシステムによる三次元再構成の評価を行う.実験では,ファントム実験により従来のフリーハンドとの再構成精度の比較をし,さらに,肢位による筋形状・筋線維変化を可視化した.
 
小野 真理子 1651030: M, 1回目発表 生体医用画像 佐藤 嘉伸
title : Automated segmentation of thin artery branches from CT images
abstract : Blood vessels are the areas which can cause fatal injury to a patient if they are erroneously cut off during operation intentionally. Therefore, it is important to show the region of the blood vessel and structure information to the doctor. It is desirable to automatically extract from the 3D-CT image by computer, and use it for image diagnosis and treatment planning. However, variations among individuals in the blood vessel region are larger than those of other organs. In the example of the abdomen region, individual differences appear in features such as position, number, branching, diameter and direction, so it is not easy to recognize their patterns. In the previous study, the search range was constrained utilizing the characteristics of the arterial branch, i.e., 1) it has a tree structure connecting an organ and the aorta, and 2) the variation of the region where the blood vessel flows into each organ is relatively small. We tried to improve the accuracy of blood vessel extraction, but thin blood vessels around 2 voxels (1 mm) diameter failed due to insufficient contrast, partial volume effect, etc. Therefore, in this research we aim to extract thin blood vessels. In this presentation, we will organize the limitations in the previous method and investigate the existing literature which addresses the problem of extracting the thin blood vessels to verify applicability to our research.
language of the presentation : Japanese
 

会場: L2

司会: Alexander Plopski
池田 祥平 1651008: M, 1回目発表 ソフトウェア工学 松本 健一
title: Toward Automatic Generation of Step to Reproduce Using Operation History
abstract: Bugs found in OSS (Open Source Software) are reported through bug tracking system managed by the OSS project. The bug tracking system also supports reporting information necessary for fixing by the bug finder. However, it is not easy for the bug finder to report the step to reproduce of bug. In this research, we automatically collect operation history (e.g. command history, file change history, method execution history etc.) worked for development. And we automatically generate the step to reproduce from the operation history.
language of the presentation: Japanese
 
音田 恭宏 1651029: M, 1回目発表 ユビキタスコンピューティングシステム 安本 慶一

title: Development of sitting posture estimation method using a chair with motion sensors

abstract: In this research, we aim to improve productivity and maintain health by estimating the physical burden and mental state correlated with the sitting posture of office workers.Continuing to sit with poor or same posture over a long time period causes burdens on the waist and shoulders, then leads to mental stress and decreases in productivity and work efficiency. Therefore, it is important to improve the sitting posture and prevent to sit with poor or same posture for a long time.In order to estimate the sitting posture, we have developed and evaluated an office chair with motion sensors attached and sitting posture estimation model.

language of the presentation: Japanese

 
梅木 寿人 1651020: M, 1回目発表 ユビキタスコンピューティングシステム 安本 慶一
title: Design of dynamic hazard map system using real-time precipitation information
abstract:Hazard maps indicate records of past disasters and evacuation information such as evacuation ranges based on predicted disaster damage, the degree of damage and evacuation routes, and sites. This information is overlaid on existing maps. However, existing hazard maps show static information of risks determined based on past disaster events or damage predictions. Therefore, users cannot understand their own disaster risk under actual hazardous situations by just looking at the existing hazard map. In this study, we propose a dynamic hazard map system that visualizes real-time changes and the expected damage and shows the necessity of evacuation to its users while advising them on disaster action.
language of the presentation: Japanese
 
北野 和哉 1651041: M, 1回目発表 光メディアインタフェース 向川 康博
title:Experiment for toward density estimation of scattering media using Time of Flight camera
abstract:Non-contact measurement of concentration is important to chemical and food area.If change concentration of liquid and illuminate light this liquid, then changed optical path length by scattering and reflection.Therefore, frequency responses are changed by concentration of liquid.However, it has not been analyzed frequency responses to change concentration of liquid until now.In this study, we propose method that estimate concentration of liquid by frequency responses using frequency variable Time of Flight camera.We measure variable concentration of liquid and confirm change frequency responses from amplitude data of time of flight camera.
language of the presentation: Japanese
 
梶原 武紘 1651032: M, 1回目発表 光メディアインタフェース 向川 康博
title: Non-rigid registration of human embryo serial-section images by blending of rigid transforms
abstract: This research proposes a novel non-rigid registration method by the blending of rigid transforms. Conventional methods estimate the displacement of each point, while the proposed method estimates rigid transform in local regions and blending them. The method is applied to the non-rigid registration of human embryo serial-section images. The 3D volume data is reconstructed from the registration result. The accuracy of the registration is evaluated by the volume data and quantification method.
language of the presentation: Japanese
 
飯塚 翔 1651005: M, 1回目発表 大規模システム管理 笠原 正治
title: On coalition formation approach for cooperative spectrum sensing in cognitive radio networks
abstract: In cognitive radio networks, it is important for secondary users (SUs) to accurately sense wireless channels for primary users (PUs). To improve the detection performance, cooperative spectrum sensing is proposed, in which SUs form distinct groups and the SUs in a group share the same sensing result. Parameters such as the miss-detection rate and false-alarm one of each SU are required to form appropriate groups, however, most of previous studies treat these parameters as known. In this study, these parameters are treated as unknown and each SU automatically learns the optimal group formation through repeated and varied attempts of group formation. Group formations are repeated with considering an exploration-exploitation tradeoff by treating this task as a multi-armed bandit problem. Performance of proposed method is evaluated through simulation experiments.
language of the presentation: Japanese
 

会場: L3

司会: 久保 尋之
川上 蓮也 1651035: M, 1回目発表 ロボティクス 小笠原 司
title: Refinment of disparity map by using Semantic Segmentation
abstract: Recently, there has been a demand for a vision sensor for quickly obtaining three-dimensional position information in fields such as automatic driving.Especially, since stereo cameras can obtain both RGB and distance informattion , they are used in various robots. However, the accuracy of distance estimation in stereo cameras is not yet sufficient. In this research, I propose a method to combine semantic segmentation to solve this problem.
language of the presentation: Japanese
発表題目: Semantic Segmentationを用いた高精度な視差マップの生成
発表概要: 近年,自動運転などの分野において3次元位置情報を素早く得るためのビジョンセンサが求められている.中でもステレオカメラは安価かつRGB情報と距離情報の両方を得ることができるため,様々なロボットで利用されている.しかし,ステレオカメラにおける距離推定の精度は未だ十分ではない.本研究では,この問題を解決するためSemantic Segmentationを組み合わせる手法を提案する.
 
井上 智章 1651016: M, 1回目発表 ロボティクス 小笠原 司
title: Person tracking method for keeping a comfortable distance between human and robot
abstract: Personal robots are expected to perform healthcare tasks and to support the housekeeping due to the increase of elderly people and solitary persons. In order to live with humans, it is necessary for the robot to keep humans comfortable by managing its behavior. In this research, we propose a tracking method for mobile robots to reduce the human's discomforting feeling caused by the robot's proximity.We describe the issue of humans' discomfort from a robot's close tracking behavior over a long time. We also introduce some related researches about the designing of robot behavior to keep distance to humans.
language of the presentation: Japanese
 
石谷 智之 1651010: M, 1回目発表 数理情報学 池田 和司
Title : Reconstruction of distance matrix from projected data of subsets
Abstract:Reconstructing relationship between processed data obtained from distinct information source is an important problem. Because getting such processed data, it is difficult to know the relationship due to the difference of their component meaning. If we treat this problem as multi-view learning, we can refer to a paper: "Latent viewpoint estimation from observations" to solve the problem. In this paper, Author assumes processed data by different view from all data are given and try to measure the distance between data by setting data space as latent space shared by all views. However, in the case, processed data obtained from distinct datasets, is not assumed. So, in my research, assuming distinct datasets is overlapping data subsets and processed data is obtained from data subset matrix of latent space by MDS and orthographic projection, I try to reconstruct data matrix by solving inverse problem of MDS and orthographic projection by using "shape and motion from Image stream under orthography". To do that, I try to find the relationship between processed data by using the data matrix of latent space.
language of the presentation: Japanese

 
阪口 栄穂 1651051: M, 1回目発表 環境知能学 萩田 紀博
title:Development of an attention guidance system by displaying small information correspond to states of the car driver
abstract:Although the number of car accidents tends to decrease, but accidents are still occurring at 500,000 cases per year in 2016. Many of accidents mainly caused by human errors of careless relate to a safe driving violation. Therefore, in this research, we propose the system to guide drivers' attention without stress for the purpose of decreasing car accidents. In particularly, the research describes the method of judging recognition based on the relationship between gazing points and the movement of the object.
language of the presentation:Japanese
発表題目:自動車運転手の状態に対応したさりげない情報提示による注意誘導システムの開発
発表概要:自動車事故の発生件数は減少傾向にあるが、平成27年の事故発生件数は年間50万件と依然多い。事故要因の内訳は運転手が注意を怠っていた安全運転義務違反が多数を占めている。そこで本研究では運転手の注意をさりげなく(運転手に負荷をかけずに)誘導するシステムを提案し、自動車交通事故の減少を目指す。特に注視点と対象の動きの関係から視認の有無を判定する方法を述べる。
 
生田 和也 1651006: M, 1回目発表 知能コミュニケーション 中村 哲
title: Neural language model capable of generating want to say for informational provision
abstract:Information provision by natural language is important for enhancing usability in information guidance systems and so on. In typical language generation, it is expressed by embedding the information that you want to convey in the prepared text template in advance. However, since the diversity of information that can be expressed depends on the number of templates, when there are not enough numbers, it may be a mechanical expression lacking naturalness. In linguistic generation using the neural language model proposed in recent years, reports that improving the naturalness of sentences without impairing the information to be presented have been made. In this research, we aim to construct a system that extends the preceding research mentioned above and enables languageization of time-series information dynamically changing regardless of specific domain / expression.
language of the presentation: Japanese
 
山田 悠貴 1651117: M, 1回目発表 ソフトウェア設計学 飯田 元

title: Support method of function inlining using cohesion

abstract: In software development, refactoring which is the work of modifying existing source code is important. In this research, we propose an approach that recommends pairs of functions to improve cohesion by inlining method refactoring.