2019 Fiscal Year Final Research Report
A study on sparse representation of high-resolution video based on directional tensor dictionary
| Project/Area Number |
17K14683
|
|---|
| Research Category |
Grant-in-Aid for Young Scientists (B)
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Communication/Network engineering
|
|---|
| Research Institution | The University of Kitakyushu |
|---|
Principal Investigator |
Kyochi Seisuke 北九州市立大学, 国際環境工学部, 准教授 (70634616)
|
|---|
| Project Period (FY) |
2017-04-01 – 2020-03-31
|
| Keywords | テンソル辞書 / スパース表現 / 高解像度映像 / 信号圧縮 / 信号復元 / 凸最適化 |
|---|
| Outline of Final Research Achievements |
In this study, a high-resolution video sparse representation method based on a directional tensor dictionary is established.In order to represent large scale tensor data such as high-definition, multi-spectral, multi-viewpoint video sparsely (for signal compression and restoration), a large tensor dictionary is required. However, due to the multidimensional nature of tensor, learning tensor dictionaries becomes difficult when the number of tensor dimensions becomes large. Since directional correlations between pixels of gray scale images (high correlations in the vertical, horizontal and diagonal directions) are also possible in video, we designed a fixed large-scale tensor dictionary with directional elements and developed an algorithm for sparse representation of high-resolution video.
|
| Free Research Field |
信号処理
|
| Academic Significance and Societal Importance of the Research Achievements |
高解像度映像スパース表現のための大規模テンソル辞書を非学習型辞書の多層線形結合によって生成するコンセプトは非常にユニークであり，またその効果を確認することができた．また本研究で開発したエピグラフ変形による各層のパラメータ学習手法は，近年活発に研究が進められている深層ニューラルネットワーク（DNN）のパラメータ学習に応用できると考えられる．DNNの性質はブラックボックスになっている部分が多く，これまで経験的な試行錯誤によって開発が進められてきたが，エピグラフ変形を架け橋として，理論が整備されている凸最適化工学の知見を活用しながら，更に高度なDNNを構築できると考えられる．
|
