Study on the application of machine learning technologies to image compression coding
| Project/Area Number |
18K11360
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 61010:Perceptual information processing-related
|
|---|
| Research Institution | Chiba Institute of Technology |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2018-04-01 – 2021-03-31
|
| Project Status |
Completed (Fiscal Year 2020)
|
| Budget Amount *help |
¥4,290,000 (Direct Cost: ¥3,300,000、Indirect Cost: ¥990,000)
Fiscal Year 2020: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Fiscal Year 2019: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Fiscal Year 2018: ¥1,690,000 (Direct Cost: ¥1,300,000、Indirect Cost: ¥390,000)
|
|---|
| Keywords | 画像符号化 / 機械学習 / 深層学習 / ディープラーニング / 画像認識 / フレーム間予測 / 画質推定 |
|---|
| Outline of Final Research Achievements |
In this research, we approached new technologies that combine machine learning and image compression coding from various perspectives. First, we proposed a new interframe prediction technology using DNN (Deep Neural Network) and a new image expression technology based on mult-class dictionaries designed by machine learning. By many experiments, it was clarified that they have better coding performance than standard image coding methods such as HEVC. We also proposed an image quality estimation method using the intermediate layer outputs of DNN, and showed that it can be used for performance evaluation of not only coding noise but also new compression methods such as texture synthesis. Furthermore, it was clarified that the singular value decomposition of the fully connected layers and the adaptive quantization for the convolution layers in DNN enable a significant reduction in the amount of information of DNN itself with almost no decrease in image recognition accuracy.
|
| Academic Significance and Societal Importance of the Research Achievements |
「画像圧縮符号化」が基本的には原画忠実性を評価基準とした技術であるのに対し，深層学習に代表される「機械学習による画像処理」は感性的な同一性や自然さに重点を置いた評価基準に基づく処理であることから，2つの技術は互いに相容れないのではないかと考えられ，機械学習を画像通信のキーテクノロジである画像符号化へ応用するアプローチは研究開始当初は本格的に行われていなかった．本研究によって，予測・変換・画質推定・深層学習のモデル圧縮など，画像符号化と機械学習が補い合うことのできる様々な要素の存在が明らかになり，今後の該分野の発展を見据えた指針を示すことができた．
|
Report
(4 results)
Research Products
(21 results)
