基於機器學習之360度視訊的 VVC快速畫面間預測演算法;Machine Learning Based Fast Inter Prediction Algorithm of VVC for 360-degree Videos

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/86346

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/86346

Title:	基於機器學習之360度視訊的 VVC快速畫面間預測演算法;Machine Learning Based Fast Inter Prediction Algorithm of VVC for 360-degree Videos
Authors:	李穎;Lee, Ying
Contributors:	通訊工程學系
Keywords:	360度視訊;EAC（equi-angular cubemap）;VVC（versatile video coding）;畫面間編碼;快速演算法;LNN（light-weighted neural network）;360-degree videos;EAC(equi-angular cubemap);VVC(versatile video coding);inter frame coding;fast algorithm;LNN（light-weighted neural network）
Date:	2021-07-19
Issue Date:	2021-12-07 12:35:53 (UTC+8)
Publisher:	國立中央大學
Abstract:	VVC（Versatile video coding）可降低高畫質視訊傳輸位元率，但VCC之編碼時間複雜度過高使其很難在即時傳輸設備上實現，也因此VVC編碼之快速演算法為視訊編碼中重要研究方向。EAC（equi-angular cubemap）格式為360度視訊格式之一，其相較於ERP（equirectangular projection）格式能減少冗餘資訊，然則現有VVC畫面間編碼的快速模式決策與深度決策演算法尚無針對EAC格式設計，因此本論文提出針對EAC格式設計之畫面間編碼快速劃分深度與模式決策演算法，其考慮EAC格式各面之影像內容相連性，與畫面間面之相關性，協助快速畫面編碼決策之準確性，並且畫面間編碼之深度決策與模式決策皆考量VVC新增之affine merge mode進行設計。又，與現有畫面間快速演算法方案相比，本論文採用LNN（light-weighted neural network）作為分類器，比經驗法則更能適應視訊內容之多樣性，並且相較於深度學習方案，僅使用中央處理器（CPU）便可以進行分類。實驗結果顯示本論文所提方案相較於VTM 7.0，平均可節省21%的編碼時間，並僅有1.03%BDBR之上升，與現有採用經驗法則之方案相比亦節省較多的編碼時間節省。;VVC (versatile video coding) can reduce the bitrate of the high-resolution videos before transmission. However, the encoding complexity of VVC is extremely high cause it hard to implements in real-time hardware. Therefore, fast algorithm of VVC encoder is important. EAC (equi-angular cubemap) format has lower redundant information than ERP format. There is not a fast inter mode or depth decision algorithm about EAC format within the survey of existing literatures. Accordingly, this paper proposed the fast inter prediction algorithm of VVC for EAC format to facilitate mode decision and depth decision process, which taking the inter prediction information of face and face boundary’s connection in EAC format into consideration. Furthermore, this paper considered method of affine merge mode which added by VVC in fast inter mode decision and depth decision. Compare with widely used classification models in VVC fast inter coding algorithms, LNN (light-weighted neural network) can better adjust oneself to different video and coding conditions than rule of thumb and just depends on CPU execution which is difficult on deep learning. Experimental results show that the proposed method reduce the encoding complexity of VTM7.0 about 21% with 1.03% BDBR (Bjontegaard delta bit rate) increasement in average and better than the rule of thumb.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	52	View/Open

社群 sharing

Loading...