2019 Fiscal Year Annual Research Report
深層学習による主観的高画質化を目指した新しい圧縮技術の研究
| Project/Area Number |
19J14620
|
|---|
| Research Institution | Waseda University |
|---|
Principal Investigator |
CHENG ZHENGXUE 早稲田大学, 理工学術院, 特別研究員(DC2)
|
|---|
| Project Period (FY) |
2019-04-25 – 2021-03-31
|
| Keywords | 画像圧縮 / 深層学習 / 高画質化 |
|---|
| Outline of Annual Research Achievements |
最近超高解像度、360度といったマルチメディアコンテンツのさらなる大容量化に伴い、より高効率な圧縮技術が求められている。令和1年度は、前述した目的を達成するために、従来の映像符号化標準に従い、提案するユーザーの感じる主観画質の高画質化を目的とし、CVPRとICIP国際学会で筆頭著者として発表を行うと共に、査読付き論文誌と令和２年度の国際学会ICASSPとCVPRへの採択が確定していた。
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
So far, the progress is conducted smoothly according to my plan. I have finished my target in the first year. And I have seperate my goals and future plan into two parts as followed.
Regarding deep learning based lossy image compression, I have reached a milestone, that is, reaching the performance of HEVC on both objective quality metrics and subjective quality, and the next plan for is to outperform VVC. Besides, some of our experiments have validate the possibility to achieve this target.
Regarding deep learning based lossy video compression, I have reached a milestone to achieve comparable performance with H.264. The next step is to outperform HEVC, or even VVC. Specially, my plan is to develop better P-frame prediction and B-frame prediction algorithms and incorporate them into inter-encoded video compression algorithms.
|
| Strategy for Future Research Activity |
To promote the development of research, we are considering the following two strategies.
The first is to promote the learned image compression. I will develop a better entropy model by extending current single Gaussian to Gaussian mixture likelihoods. Second, I will enhance the network structure by consider some novel modules such as attention. The CVPR, a flagship meeting in the field of computer vision will have the event "Workshop and Challenge on Learned Image Compression” (CLIC) and I will participate in the challenge. Not only I can compare my results with research teams from other universities and companies, but also I could compare with conventional compression standards, because some hybrid methods are also submitted to this challenge.
The second is to enhance the performance of learned video compression. My plan is to enhance the overall performance module-by-module. Then what I can do include the better video prediction (P-frames) algorithms, better video interpolation algorithms (B-frames) and better network to encode residual and motion vectors etc. Based on the knowledge of learned image compression, I can extend the entropy models to videos and combine all these modules together. Finally, I can achieve better results than my current result.
Finally, through this research and development, we can contribute to the further development of compression technology.
|
