基於效率化ETH-CNN技術來快速切割CU的方法

Fast HEVC CU partition based on lightened ETH-CNN and Rich CPH datasets

鐘珮瑀(Pei-Yu Zhong)

指導教授 ： 王家祥

國立清華大學/電機資訊學院/資訊工程學系所/碩士(2018年)

摘要

H.265 (或稱HEVC)是2013 年ISO通過的的視訊與影像壓縮技術標準。 H.265 壓縮的概念是將一張畫面切成數個Coding Unit(CU)以z-scan 的方式依序作壓縮；在H.265 的標準下，提出了Coding Tree 的概念。Coding Tree 是以Coding Unit, Predict Unit, Transform Unit 所組成，為了提高壓縮後影像的品質，將64x64 的CU 再細分成32x32, 16x16, 8x8 總共四層。除此之外，H.265 也提出多項新的機制，目的是增加壓縮的影像品質和降低儲存所需要的位元；但伴隨而來的是複雜的運算以及大量的運算時間需求。
由於其中的CTU partition對HEVC中的編碼複雜度影響最大，因此大多的方法會通過簡化CTU的處理來降低複雜度，本論文研究探討運用Convolutional Neural Network(CNN)技術來獲取有效率的加速效果，達到編碼速度的提升。
在以前的一些方法中，必須手動提取一些特徵，如RD成本，量化參數（QP）和紋理複雜度來預測CTU partition，這些特徵依賴與CTU partition結果之間關係的先前知識，因此本論文藉由改進ETH-CNN獲取特徵以及利用有大量數據的CPH-Intra database，可以自動學習CTU partition結構的特徵來獲取有效率的加速效果。

關鍵字

高效率視訊編碼 ； 編碼樹分割 ； 卷積神經網路

並列摘要

HEVC is intended to provide significantly better coding efficiency than H.264/ AVC and its predecessors, but it increase the expense of extremely high encoding complexity. In particular, in HEVC, a quad-tree partition of the coding unit (CU) which is consumes a large proportion of the encoding complexity, lead to the exhaustively search for the best rate-distortion optimization (RDO) partition. In [1], a deep learning approach (convolutional neural network, ETH-CNN) to predict the CU partition for reducing the HEVC complexity at intra-modes was proposed. Their CU partition scheme is considered for resolving the entire coding tree unit instantaneously instead of one level at a time. Thus, a large-scale training dataset including substantial CU partition data is necessary for solving this complicated problem.
In this thesis, a lightened ETH-CNN, which augmenting ETH-CNN model through some useful CNN ideas, such as Network-in-Network [2], Maxout Networks [3], Batch Normalization [4] etc. to improve the prediction accuracy plus reduce the computational complexity as well. The experimental results demonstrate that the lightened approach provides an increase accuracy (64x64 to 32x32) in the CU partition prediction, a decrease (in QP22) in the CU partition time.

並列關鍵字

HEVC ； Convolutional Neural Networks ； Network-in-Network ； Maxout Networks ； Batch Normalization

參考文獻

[1]. Mai Xu, Tianyi Li, Zulin Wang, Xin Deng, Ren Yang and Zhenyu Guan, “Reducing Complexity of HEVC: A Deep Learning Approach,” IEEE Transactions on Image Processing, Vol. 27, No. 10, pp. 5044-5059, Oct. 2018.

[2]. Min Lin, Qiang Chen, Shuicheng Yan, “Network in Network,” arXiv: 1312.4400, Mar 2014.

[3]. Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio, “Maxout Networks,” Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 Atlanta, USA, June 16 - 21, 2013.

[4]. Sergey Ioffe and Christian Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” arXiv: 1502.03167, Mar 2015.

[5]. Tianyi Li, Mai Xu and Xin Deng, “A Deep Convolutional Neural Network Approach for Complexity Reduction On Intra-Mode HEVC,” in 2017 IEEE International Conference on Multimedia and Expo (ICME), Hong Kong, 10-14 July 2017, pp. 1255-1260.