採用資料探勘之快速運動估測演算法與DSP實現

高效能視訊編碼(high efficiency video coding: HEVC，又稱H265)是最新一代的視訊編碼標準，HEVC採用更彈性的視訊編碼模組，分別為編碼單位(coding unit: CU)、預測單位(prediction unit: PU)和轉換單位(transform unit: TU)，其中PU採用運動估測(motion estimation: ME)來進行畫面間預測(inter prediction)，為了提高PU預測的準確性，HEVC允許ME模組執行多重參考畫面(multiple reference frame: MRF)來進行更精準的預測，這也導致ME模組需要更龐大的計算量，以至於難以達到即時視訊傳輸之應用。　　為了降低HEVC中ME模組的計算複雜度，最近Yang等學者提出了快速參考畫面選擇演算法[4]，他們經由統計發現在MRF的先進運動向量預測(advanced motion vector prediction: AMVP)中，最小位元率-失真成本(rate-distortion cost: RD cost)所對應的參考畫面與最佳參考畫面間存在很高的相關性。因此，Yang提出當HEVC在進行ME模組之前，利用預先設定的臨界值(threshold)來判斷AMVP中最小RD cost所對應的參考畫面是否為最佳畫面，來加速ME模組的運算。由於Yang主要是根據AMVP在MRF間的相關性統計，並藉由量化參數(quantization parameter: QP)和PU大小來設定臨界值，然而當畫面場景變化較大或背景特徵較複雜時，將導致所需平均參考畫面變多，以至於無法大幅提升ME模組的速度。　　當HEVC應用在4K超高解析度(ultra high definition: UHD)視訊編碼時，ME模組將面臨大數據(big data)的資料量。由於MRF間具有很高的時空關聯性(temporal-spatial correlation)，為了改善Yang等學者的方法，本論文利用資料探勘(data mining)的技術，在PU中分別找尋出適合ME模組的屬性(attributes)，並在編碼過程中提取出每個屬性相對的數據，接著將數據以ARFF (attribute-relation file format)的檔案格式輸入至軟體工具WEKA[9]，透過C4.5機器學習演算法[10]來訓練出適合的最佳參考畫面選擇的決策樹(decision tree)，最後將此決策樹運用在PU結構中，來快速選擇最佳的參考畫面，大幅降低MRF的計算複雜度，加速ME模組的速度。　　此外，本論文採用ADSP-BF609開發板來完成所提出快速ME模組的DSP實現。首先，對於DSP內部進行記憶體配置的最佳化，我們將運算複雜度高的PU模組從L3配置到L2中，提高PU模組的執行效率；接著，使用ADSP-BF609開發板專用指令，將原程式碼進行修改與優化，並進行實驗模擬與分析。最後，由實驗結果得知，本論文所提出的方法當參考畫面為4張(MRF=4)時，與HM16.7[]和Yang的時間改善率(time improving ratio: TIR)相比，TIRHM和TIRYang分別為70.09%和46.03%，當MRF=8時，TIRHM和TIRYang分別為82.49%和35.04%。本論文所提出的方法，除了能加速HEVC編碼過程外，更可以得到與HM16.7差異不大的影像品質。

關鍵字

none

並列摘要

High efficiency video coding (HEVC) is the newest video encoding standard. HEVC adopts some new coding structures including coding unit (CU), prediction unit (PU) and transform unit (TU). In the PU structure, HEVC adopts motion estimation (ME) module to achieve inter prediction. In order to improve the accuracy of PU prediction, HEVC allows the ME module performing on multiple reference frames (MRF). Although the ME-MRF can enhance the PU performance and allow the encoder to search a better reference frame from several previous pictures, the computational complexity of the ME-MRF dramatically increases. Thus, the real-time applications of HEVC will be limited. 　　In order to reduce the computational complexity of ME module in HEVC, Yang et al. proposed a HEVC fast reference picture selection recently [4]. After the statistical analysis of ME-MRF, they found that a high correlation exists between the best reference frame and lowest rate-distortion cost (RD cost) associated with advanced motion vector prediction (AMVP). Therefore, they use the predefined threshold to determine whether the AMVP-selected reference frame is the best reference frame. However, the predefined threshold is inefficient when the video sequence with active motion and complicate background. 　　To further improve the accuracies of MRF prediction, we propose a fast ME algorithm based on data mining to speed up encoding process of HEVC encoder. Firstly, we find that there is a high temporal-spatial correlation existing in the MRF. Then, we find appropriate attributes from ME module and extract the corresponding data. We save these attributes as an attribute-relation file format (ARFF). And then, the ARFF is performed on WEKA [9] to train the ME-MRF decision trees using the C4.5 algorithm [10]. Finally, we employ the created decision tree to achieve fast selection of the reference frames. 　　In addition, to further achieve the DSP realization for the proposed fast motion estimation algorithm, we embed the codec on the ADSP-BF609. We re-allocate the function of consuming module from L3 DDR-RAM to L1 and L2 SRAM to speed up the encoding time of HEVC. Simulation results show that the proposed method can achieve average time improving ratio (TIR) with MRF=4 about TIRHM=70.09% and TIRYang= 46.03% when compared to HEVC (HM16.7) [7] and Yang’s method, respectively. The gains of TIRHM=70.09% and TIRYang= 46.03% with MRF=8 can be reached. It is clear that the proposed method can efficiently increase the speed of HEVC encoder with insignificant loss of image quality.

並列關鍵字

none

參考文獻

[17] 吳世欣編著“應用於DSP之高效率記憶體配置來實現H.264視訊編碼器” , 義守大學電子工程學系論文, 中華民國一百零三年七月

[4] S. H. Yang and K. S. Huang, “HEVC fast reference picture selection,” Electronics letters 10th December 2015 Vol. 51 No. 25 pp. 2109–2111

[6] High efficiency video coding, docment ITU-T Rec. H.265, Oct. 2014.

[10] J. R. Quinlan, “C4.5: Programs for Machine Learning,” Morgan Kaufmann Publishers, 1993.

[15] Z Pan, P Jin, J Lei, Y Zhang, X Sun and S Kwong “Fast reference frame selection based on content similarity for low complexity HEVC encoder,” Journal of Visual Communication and Image Representation 40, pp.516-524, Oct. 2016

國際替代計量

採用資料探勘之快速運動估測演算法與DSP實現

全文下載

主題瀏覽