Acceleration of Kvazaar HEVC Intra Encoder With Machine Learning
Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Scientific › peer-review
Details
| Original language | English |
|---|---|
| Title of host publication | 2019 IEEE International Conference on Image Processing (ICIP) |
| Publisher | IEEE |
| Pages | 2676-2680 |
| Number of pages | 5 |
| ISBN (Electronic) | 978-1-5386-6249-6 |
| ISBN (Print) | 978-1-5386-6250-2 |
| DOIs | |
| Publication status | Published - 26 Aug 2019 |
| Publication type | A4 Article in a conference publication |
| Event | IEEE International Conference on Image Processing - Duration: 1 Jan 1900 → … |
Publication series
| Name | IEEE International Conference on Image Processing |
|---|---|
| ISSN (Print) | 1522-4880 |
| ISSN (Electronic) | 2381-8549 |
Conference
| Conference | IEEE International Conference on Image Processing |
|---|---|
| Period | 1/01/00 → … |
Abstract
complexity of High Efficiency Video Coding (HEVC) poses a real challenge to HEVC encoder implementations. Particularly, the complexity stems from the HEVC quad-tree structure that also has an integral part in HEVC coding efficiency. This paper presents a Machine Learning (ML) based technique for pruning the HEVC quad-tree without deteriorating coding gain. We show how ML decision trees can be used to predict a depth interval for a quad-tree before the Rate-Distortion Optimization (RDO). This approach limits the number of RDO candidates and thus speeds up encoding. The proposed technique works particularly well with high-quality video coding and it is shown to accelerate the veryslow preset of practical Kvazaar HEVC intra encoder by 1.35× with 0.49% bit rate increase. Compared with the corresponding preset of ×265 encoder, Kvazaar is 2.12× as fast at a cost of under 1.21% bit rate overhead. These results indicate that the optimized Kvazaar is the leading open-source encoder in high-quality HEVC intra coding.
Publication forum classification
Field of science, Statistics Finland
Downloads statistics
No data available