| 摘要: | 現今科技日新月異,科技與人們的生活息息相關,但隨著數位內容需求的增加,尤其是在高解析度(如4K、8K甚至更高)的視頻領域,人們對於影像品質和視覺體驗的期望也越來越高,因此H.266/VVC採用許多先進的技術來解決人類的需求,隨之而來的影響是編碼計算複雜度的提升。
本論文運用機器學習跟深度學習的方式,將其應用在H.266/VVC畫面間預測,首先運用機器學習中的支持向量機(SVM)將CU做第一次劃分,接著運用卷積神經網路(CNN)將CU做第二次劃分,針對H.266/VVC新增的多類型樹劃分模式,採用隨機森林分類器(RFC)將CU劃分的更精細,最後再結合CU-PU Decision進行第三次的劃分,透過漸進式的劃分使得編碼過程,不需再計算冗長的碼率失真代價函數。第四章則探討將隨機森林分類器,銜接在不同位置上,所造成的影像品質與時間節省的差異。最終實驗結果與H.266/VVC相比,平均整體 BDBR 為2.50百分比,而編解碼時間可以達到64.67百分比的節省。;In today’s rapidly evolving technological era, people′s lives are closely connected to technology. With the increasing demand for digital content, especially in high-resolution video(such as 4K, 8K,and even higher),expectations for image quality and visual experience are rising. Therefore, H.266/VVC employs many effective techniques to address these needs.
However,this comes with the consequence of increased encoding complexity.This paper applies machine learning and deep learning techniques to inter-frame prediction in H.266/VVC. First, the Support Vector Machine (SVM) from machine learning is used for the initial division of Coding Units (CUs). Next, a Convolutional Neural Network (CNN) is
applied for a second division of the CUs. For the new multi-type tree partitioning mode introduced by VVC, a Random Forest Classifier (RFC) is used to further refine the CU division.
Finally, the CU-PU Decision is combined for a third division. Through this progressive division, the encoding process avoids the need for calculating the lengthy rate-distortion cost
function. Chapter four explores the impact on image quality and time savings when connecting the Random Forest Classifier at different stages. The final experimental results show that,
compared to H.266/VVC, the average Bjøntegaard Delta Bit Rate (BDBR) is 2.50%, and encoding/decoding time is 64.67%. |
