用於HEVC畫面間視訊編碼的四元樹樣版匹配

Abstract

視訊壓縮技術與標準日新月異，為了提高壓縮效率，新的方法陸續被提出，用於畫面間估測的樣版匹配(Template Matching)為其中一個研究分支。自樣版匹配的概念被提出以來，相關的研究持續進行，例如將多個樣版匹配的結果做加權平均以降低白雜訊(white noise)、混合區塊匹配(Block Matching)的結果來降低估測誤差，或是進一步將樣版匹版以Matrix Completion的問題來處理，以提高估測的準確性。以上所提的研究成果皆能進一步提高壓縮率，然而不同方法間存在一定程度的衝突，多數無法混用，因此對壓縮率的改善無法達到加乘的效果。 本論文所提出的四元樹樣版匹配(Quadtree Template Matching)是將樣版匹配的概念進行擴充，這使得傳統樣版匹配僅是四元樹樣版匹配的一個特例，反過來說，當四元樹樣版匹配的最大葉節點深度為零時，則退化成傳統樣版匹配。而過去基於傳統樣版匹配的研究成果，可套用於四元樹樣版匹配，且在相同影像品質下的壓縮率能得到更進一步的提升、達到加乘的效果。由本論文在相同參數下所做的四項實驗共24組數據可以發現，四元樹樣版匹配在壓縮率的改進更優於傳統樣版匹配，且藉由樣版匹配方法本身的改進，所對應的四元樹樣版匹配版本的壓縮率提升會更為顯著，例如由本論文其中兩組實驗數據顯示，當不同樣版匹配方法為BD-rate位元率帶來-0.48%與-1.08%的改善時，四元樹樣版匹配所對應的改善分別為-0.81%與-1.50%。 本論文同時探討樣版匹配對merge mode與AMVP的影響，並提出兩個方法，其中一個方法可以進一步提高壓縮率，另一個方法可用以做為衡量的依據。

With the ever-changing video compression technology, many studies have been proposed to improve the compression efficiency, in which the template matching for inter-frame prediction is one of the research fields. Since the template matching concept was introduced, many relevant studies have been proposed. For example, the weighted average of multiple template-matching predictors can reduce white noises; combining the predictor of block matching can lower the prediction error; furthermore, it can improve the prediction accuracy by treating template matching as a matrix completion problem. Many template matching related methods can increase the compression rate, but there is a certain degree of conflict between them. Most of them cannot be combined to increase the compression rate even further, so the best one of all methods limits the performance of template matching. This paper introduces a method, Quadtree Template Matching (QTM), which expands the notion of template matching and makes traditional template matching a special case of QTM, i.e., when the maximum quadtree depth is zero, the QTM degenerates into a traditional template matching. QTM can be applied to prior researches that are based on traditional template matching, and compression rate can be improved further under the same image quality. According to 24 sets of data in this paper, QTM outperforms traditional template matching in compression rate, and through the improvement of template-matching extended methods, the compression performance of corresponding QTM extensions would be more significant. For example, experimental data in this paper show that when the BD-rate bitrate of two different template-matching methods were -0.48% and -1.08% improved, the BD-rates of the corresponding QTM extensions would be -0.81% and -1.5%. This paper also discusses the impact of template matching to merge mode and AMVP, and proposes two methods; one can further improve the compression rate, and another can be used as basis for measurement.