博
士
学
位
論
文
Doctoral Dissertation内容の要旨 及び
審査結果の要旨
Dissertation Abstract and
Summary of the Dissertation Review Result
第
35号
The Thirty-Five Issue
2020年3月 March, 2020 The University of Aizu
はしがき
博士の学位を授与したので、学位規則（昭和２８年４月１日文部省令第９号）第８条の規定 に基づき、その論文の内容の要旨及び論文審査の結果の要旨をここに公表する。
学位記番号に付した｢甲｣は学位規則第４条第１項（いわゆる課程博士）によるものであるこ とを示す。
Preface
On granting the Doctoral Degree to the individuals mentioned below, abstracts of their theses and the theses review results are herewith publicly announced, in according to the provisions provided for in Article 8 of the Ruling of Degrees (Ministry Of Education Ordinance No.9, enacted on April 1, 1953)
The Chinese character, “甲”, at the beginning of the diploma number represents that an individual has been granted the degree in accordance with the provisions provided for in Paragraph 4-1 of the Ruling Of Degrees (what is called “Katei Hakase,” or the Doctoral Degree granted by the University at which the grantee was enrolled.).
- 1 -
目 次 Contents
掲載順
Order
学位記番号 Diploma No.
学位 Degree
氏名 Name
論文題目 Dissertation Title
頁 Page
1
甲CI博 第79号
博士（コンピュー タ理工学）
The Degree of Doctor of Science and
Engineering
HAYASHI, Kensaku 林 憲作
自然な海底地形と人工的な海底地形 を結合したマルチスケール津波シミュ レーションのためのアルゴリズムとソフ トウェア
Algorithms and Software for Multi-scale Tsunami Simulation Combining Natural and Artificial Bottom Topography
2
2
甲CI博 第80号
博士（コンピュー タ理工学）
The Degree of Doctor of Science and
Engineering
TAN, Benying 譚 本英
非凸正則化によるスパース表現とそ の効率的な方法
Sparse Representation with Nonconvex Regularization and Its Efficient Methods
5
- 2 - Name
氏名
HAYASHI, Kensaku 林 憲作
The relevant degree 学位の種類
Doctoral degree (in Computer Science and Engineering) 博士（コンピュータ理工学）
Number of the diploma of the Doctoral Degree 学位記番号
甲CI博第79号
The Date of Conferment 学位授与日
March 19, 2020 2020年3月19日 Requirements for Degree Conferment
学位授与の要件
Please refer to the article five of “University Regulation on University Degrees”
会津大学学位規程 第5条該当 Dissertation Title
論文題目
Algorithms and Software for Multi-scale Tsunami Simulation Combining Natural and Artificial Bottom Topography
自然な海底地形と人工的な海底地形を結合したマルチスケー ル津波シミュレーションのためのアルゴリズムとソフトウェア Dissertation Review Committee Members
論文審査委員
The University of Aizu, Prof. VAZHENIN, A. P.
(Chief Referee)
The University of Aizu, Prof. NAKASATO, N.
The University of Aizu, Senior Associate Prof. YOSHIOKA, R.
The University of Aizu,Senior Associate Prof. WATANOBE, Y.
会津大学教授 アレクサンンダー ヴァジェニン（主査）
会津大学教授 中里 直人
会津大学上級准教授 吉岡 廉太郎 会津大学上級准教授 渡部 有隆
- 3 -
Abstract
The big dangerous and negative effects caused by tsunamis formulate many challenges oriented to reduce their disastrous effects. Accordingly, it is possible formulate two main issues related to mentioned above. The first issue is addressed to designing the real-time tsunami warning systems allowing to predict the basic parameters of tsunami waves based on time. The second issue is related to the long-term hazard assessment and tsunami inundation mapping. It is required to support preparedness measures and effective disaster reduction. Both approaches to the numerical modeling of tsunami wave propagation consider it a computationally intensive problem requiring the acceleration of calculations through parallel processing. In this dissertation, we developed a universal high-speed computational scheme for tsunami modeling that can used for solving both mentioned issues.
According to the real-time tsunami warning system, the mathematical model was designed for simulating wave propagation as well as the nested multi-grid algorithm for computing tsunami propagation from the initial source to the coastline involving scale switching. The originality of the proposed scheme is that digital bathymetry can be taken or developed using different sources.
Moreover, the proposed algorithm allows integrating computational grids with non-proportional steps.
Based on this algorithm, the streamed Tsunami Modeling Infrastructure was designed supporting high-speed tsunami modeling on a system with rather limited computational resources. The streamed scheme is realized by distributing bathymetries over computational resources and synchronizing of calculation for each area using buffering of boundaries between areas. Mixing CPU- based and GPU-based computations it was shown that developed system can reduce the computational burden of tsunami modeling.
The approach presented here allows for the creation of flexible and reconfigurable computational schemes with a variable set of modeling zones by sharing these data with other chains implementing modeling for other embedded coastal areas. The results presented confirm the possibility of implementing high-speed computations concurrently at the laboratory level using distributed computing in combination with CUDA- accelerators. This makes it possible to significantly improve the total calculation speed compatible with calculations on supercomputers.
According to the long-term hazard assessment and tsunami inundation mapping, the important part of the tsunami research is focused on studying the considerable influence of natural geographical objects, like islands and near-coastal bathymetry, on tsunami waves. The flexibility of the system makes possible to provide these experiments. Complementing the physical modeling, we adopted the presented system for computer simulations of crucial coastal areas. The Bathymetry and Tsunami Source Data Editor is a basic system tool for editing bathymetric and tsunami source data by including/removing artificial seawalls and submerged barriers having different shapes and sizes.
Results of numerical experiments are presented for the gridded hybrid bathymetry for several coastal areas of Japan. This makes possible to control the tsunami wave height by underwater artificial objects as well as provide studying features of the natural bathymetry. This system can help to issue recommendations for better protection of some crucial objects on a coastline.
- 4 -
Summary of the Dissertation Review Result
The dissertation is devoted for creating the algorithms and software allowing to help in solution of two issues: real-time tsunami warnings and long-term hazard assessment. Accordingly, a distributed Tsunami Modeling Infrastructure was developed to support high-speed tsunami modeling on systems with limited computational resources consisting, usually, of several personal computers supported by standard accelerators like GPU and FPGA boards. In the framework described, the following contributions were obtained.
In the first contribution, the grid-switching nested algorithm was developed for modeling tsunami propagations by carrying out a sequence of non-proportional grids with various resolutions where one is embedded into another.
According to the second contribution, the distributed streaming computational scheme was proposed and investigated allowing flexible reconfiguration of heterogeneous computing resources with a variable set of modeling zones. The scheme is based on the special synchronization protocol for data exchange between the system’s components allowing boundaries transition for each time step.
Finally, the adaptation this system to numerical tsunami simulations was implemented for crucial coastal areas supporting so-called hybrid bathymetry combining natural and artificial underwater objects. To realize this scheme the original Tsunami Source and Bathymetry editor was developed that allows tuning/editing bathymetric and tsunami source data by including/removing artificial barriers as well as specifying their placement, shapes and sizes.
The committee evaluates the significance of the dissertation by reviewing the candidate's doctoral thesis and by listening to his final presentation. The review committee’s judgment was made after the candidate made a presentation for about 50 minutes. During his presentation, the candidate answered all questions and addressed the points recommended during his preliminary review by the committee.
During the questions and answers session, the candidate answered most of questions well asked by the review committee. Final judgment was made after a discussion between all committee members in a closed discussion. The committee determines that the body of work accomplished by the candidate is relevant and important to the scientific community. The review committee judges that the dissertation has enough contributions and results and is recognized as qualified for conferment for a Doctor degree.
- 5 - Name
氏名
TAN, Benying 譚 本英 The relevant degree
学位の種類
Doctoral degree (in Computer Science and Engineering) 博士（コンピュータ理工学）
Number of the diploma of the Doctoral Degree 学位記番号
甲CI博第80号
The Date of Conferment 学位授与日
March 19, 2020 2020年3月19日 Requirements for Degree Conferment
学位授与の要件
Please refer to the article five of “University Regulation on University Degrees”
会津大学学位規程 第5条該当 Dissertation Title
論文題目
Sparse Representation with Nonconvex Regularization and Its Efficient Methods
非凸正則化によるスパース表現とその効率的な方法 Dissertation Review Committee Members
論文審査委員
The University of Aizu, Associate Prof. LI, X.(Chief Referee)
The University of Aizu, Prof. MORI, K.
The University of Aizu, Senior Associate Prof. HAMADA, M.
The University of Aizu, Associate Prof. PEI, Y.
Guilin University of Electronic Technology, Prof. DING, S.
会津大学准教授 李 想（主査）
会津大学教授 森 和好
会津大学上級准教授 モハメド ハマダ 会津大学准教授 裴 岩
桂林電子科技大学教授 丁 数学
- 6 -
Abstract
Sparse representation was viewed as a powerful model for representing various signals. The sparse model has extensive applications in real-world, not only the area of signal processing. Sparse representation aims to model a signal by a linear combination of atoms. The set of atoms are called dictionary, which is column redundant, i.e., overcomplete. For sparse representation, the choice of the dictionary is a key problem. A well-defined dictionary is a benefit for the performance of sparse representation. The most potential method is dictionary learning, which is to learn a dictionary from a set of training data sampled from the processed data or signals. The learned dictionary can closely match the inner structure of corresponding data to be processed. Therefore, the learned dictionary is adaptive for various applications. Due to the importance of dictionary learning, we have begun the study of dictionary learning to perform the sparse representation. The sparse regularization and optimization methods have played a major role in the developing of sparse representation. Currently, a mainstream is using nonconvex sparse regularization to deduce the sparsity and using effective optimization methods to solve the formulated objective functions. Generally, nonconvex sparse regularization can result in more accurate estimation of high-amplitude components than convex sparse regularization. Our research is started by considering these two aspects.
Throughout the whole thesis, the nonconvex nonseparable sparse regularizations are used to deduce sparsity for promoting the sparsity and accuracy of solutions. The determinant- type sparse measure was studied in the synthesis model and analysis model for nonnegative sparse representation. The generalized minimax-concave (GMC) sparse regularization is employed to maintain the convexity of formulated objective functions for obtaining global optima. Note that the difference of convex functions (DC) programming and DC algorithm was continuously used to solve the nonconvex minimization problems in this thesis as its power for nonconvex sparse optimization.
At first, we consider a nonnegative dictionary learning problem with the determinant-type sparsity measure, which can be viewed as a nonconvex sparse regularization (NSR). Different from most of the nonconvex sparse regularizations, the determinant sparse regularization is non- separable and can reflect the spatial and two dimensions characteristics of signals. This feature makes the determinant-based algorithms having advantages in image and video processing be- cause images and videos can be viewed as having spatial characteristics. Besides, we use the difference of convex functions (DC) programming and DC algorithm to effectively solve the formulated nonconvex objective functions. We construct synthetic data experiments and real-world data experiments to verify the effectiveness of the proposed dictionary learning algorithms. The proposed nonnegative dictionary learning algorithm is robust for recovering dictionary. Mean- while, as an expanding, the determinant-type sparse measure and DC programming are used in nonnegative sparse representation with the analysis model.
Then, as an improvement, we introduce the nonconvex generalized minimax-concave (GMC) sparse regularization to formulate dictionary learning problem. The family of NSR is used to promote the accuracy and the sparsity of solutions. However, the NSR usually leads the nonconvex objective functions. The optimization of nonconvex objective functions is challenging. The GMC sparse
- 7 -
regularization can maintain the convexity of the formulated objective functions so that the convex optimization method can be used to solve them whereby for obtaining global optima. In the sparse coding stage, we use the forward-backward splitting (FBS) method to calculate the sparse coefficients.
In the dictionary update stage, we use the DC programming and DCA to update the dictionary.
Besides, we have discussed the accelerated strategy to improve the convergence speed of the FBS algorithm. Adaptive threshold strategy is used to adjust the regularization parameters. By simulation experiments and real-world data experiments, we verify the effectiveness of the proposed dictionary learning algorithms.
What is more, we conduct a sparse coding modelling to extract key frames from videos as a practical application. Extracting key frames from a video can reduce redundancies in continuous scenes and concisely represent the entire video. For reducing the data dimension, we first use the deep learning method to extract the features of video frames as pre-processing. Then, we use the extracted features as the dictionary to formulate sparse coding. The determinant-type sparse measure is used to deduce the sparsity and DCA is used as the solver. By measuring the F-measure and summarized length, we evaluate the performance of the proposed key frame ex- traction method on a so-called Summe data set. Our proposed method is effective and performs well compared with the state-of-the-art frame extraction methods.
In short, this thesis concentrates on the subject of dictionary learning for sparse representation of signals. The main methods are in one to introduce nonconvex nonseparable sparse regularization to deduce sparsity in dictionary learning formulation, and then draw support the practical optimization approaches to handle the nonconvex formulations, so as to keep the designed dictionary learning algorithms robust and perform well.
Summary of the Dissertation Review Result
The committee members all agreed that the candidate Benying Tan passed his doctoral thesis defense with good quality in both publication record and oral presentation.
The committee members confirmed the contributions for the research in the signal processing topic about sparse representation and dictionary learning. Their concerns were proposed in Q & A unit, and the candidate provided a proper response. The promising work that introducing determinant-type sparse constraint and Difference of Convex function (DC) to create new algorithms which improve the performance in both numerical and practical experiments were also reviewed. The application for video key-frame extraction obtained a wide discussion and an expectation was generated for the future studies. Moreover, the candidate was humble but active for the challenging questions and replied with considerable points from the scope of his research.
In the dissertation, the DC programming and DC Algorithm are used to solve nonconvex optimization efficiently. The proposed dictionary learning algorithms and key frame extraction method are proven to have better performance in dictionary recovery and key frame extraction, respectively. Three main contributions of the dissertation are presented during the final review, and they are list as follows.
- 8 -
1. The candidate has devoted to the design of the dictionary learning algorithm to learn the data-adaptive dictionary with the determinant-type sparse measure for nonnegative sparse representation with synthesis model and analysis model.
2. As the improvement, the candidate proposed to use the generalized minimax-concave (GMC) penalty as sparse regularization to formulate dictionary learning model to promote the performance. Numerical experiments and real-world data verify the performance of the proposed algorithms.
3. A novel key frame extraction modeling based on the determinant-type sparse measure and the nonnegative sparse representation framework was proposed for key frame extraction from videos.
- 9 -
博 士 学 位 論 文 Doctoral Dissertation
内容の要旨 及び 審査結果の要旨 Dissertation Abstract
and
Summary of the Dissertation Review Result
第35号 The Thirty-Five Issue
2020年3月 March, 2020
発行 会津大学
〒965-8580 福島県会津若松市一箕町鶴賀 TEL: 0242-37-2600
FAX: 0242-37-2526 THE UNIVERSITY OF AIZU Tsuruga, Ikki-machi Aizu-Wakamatsu City
Fukushima, 965-8580 Japan
