[ieee 2013 ieee 17th international symposium on consumer electronics (isce) - hsinchu city, taiwan...

2
Abstract— JCT-VC joined by ITU-T VCEG and ISO/IEC MPEG is developing a new video coding standard called HEVC to achieve better coding efficiency for the emerging HD or Ultra HD video content. One of the new coding features in HEVC is increasing intra prediction modes from 9 in H.264/AVC to 35. To reduce the computation time, this paper presents a fast intra mode decision algorithm based on direction energy distribution. The saving of computation time for the first rough mode decision can be up to 68% under the negligible PSNR drop and bit rate increase. I. INTRODUCTION With the presence of HD and Ultra HD video content, Joint Collaborative Team on Video Coding (JCT-VC) develops a new video coding standard called High Efficiency Video Coding (HEVC) [1]. HEVC provides 33.7% better coding efficiency than the state-of-the-art video codec H.264/AVC by adopting a lots of new features [2]. One of these new features is increasing the number of intra prediction modes up to 35 including Planar mode, DC mode, and 33 angular modes as shown in Fig. 1 (a). However, to find the optimal mode, exhaustive search on these 35 prediction modes leads to much more computation time. Several fast intra mode decision algorithms are proposed to reduce the computation time of the encoder. The Piao’s algorithm adopted in HM1.0 which is the HEVC reference software first determines the N best candidate modes by testing all 35 modes with sum of absolute Hadamard transformed difference (SATD) [3]. Then the rate-distortion optimization (RDO) on the N candidate modes are performed to find the best intra mode. In addition to choosing N candidate modes, Zhao’s algorithm adopted in HM2.0 adds the most probable mode (MPM) selected from the neighboring prediction units (PU) to improve the coding quality [4]. The Yan’s algorithm reduces the number of candidate modes from 3 to 1 for the PU size bigger than 8x8 by exploring the correlation of the selected candidates in the first rough decision [5]. Although the number of candidate modes for RDO process are reduced, it still requires much time to test all 35 modes during the first rough decision. Silva exploits edge information of the image to speed up the first rough decision [6]. It first calculates one non-directional edge strength and four directional edge strength: horizontal, vertical, 45 0 , and 135 0 on a 4x4 PU. Then only the strongest edge orientation and its fixed eight adjacent directions are chosen for the first rough decision. This paper presents a fast intra mode decision algorithm based on direction energy distribution. First, the four main direction energy at degree 0, 45, 90, and 135 are calculated. Then construct a set according to the correlation of energy strength in the four directions. Only 11 or 13 candidate modes rather than 35 are selected for the first rough decision dependent on the direction distribution in the set. The experimental result shows that up to 68% intra mode computation complexity is reduced with 0.04db BD-PSNR degradation and 0.65% BD-rate increase. (a) (b) Fig. 1. (a) Intra prediction modes in HEVC (b) Intra mode decision algorithm in HM6.1 The rest of this paper is organized as follows. In Section 2, we present the proposed fast intra mode decision algorithm. Experimental results are shown in Section 3. Finally, this paper is concluded in Section 4. II. PROPOSED FAST INTRA MODE DECISION ALGORITHM Fig. 1 (b) describes the intra mode decision algorithm in HM6.1. To avoid exhaustively checking all intra modes in the first rough decision, this paper proposes a coarse-to-fine intra mode decision algorithm. In the first stage, it performs the coarse mode decision based on the direction energy distribution described below: Step 1. Calculate four main direction energy at degree 0, 45, 90 and 135 like [7]. An NxN-size PU is equally divided into four N/2xN/2-size parts and an N/2xN/2-size central part. The energy of a part is calculated by accumulating N/2*N/2 pixels. , , , in a NxN PU 1 Fast Intra Mode Decision for HEVC Based on Direction Energy Distribution Ce-Min Fang, Yuan-Teng Chang, Wen-Hao Chung Industrial Technology Research Institute, Taiwan R.O.C. A B C D F N N 978-1-4673-6199-6/13/$31.00 ©2013 IEEE 2013 IEEE 17th International Symposium on Consumer Electronics (ISCE) 61

Upload: wen-hao

Post on 14-Dec-2016

217 views

Category:

Documents


5 download

TRANSCRIPT

Abstract— JCT-VC joined by ITU-T VCEG and ISO/IEC

MPEG is developing a new video coding standard called HEVC to achieve better coding efficiency for the emerging HD or Ultra HD video content. One of the new coding features in HEVC is increasing intra prediction modes from 9 in H.264/AVC to 35. To reduce the computation time, this paper presents a fast intra mode decision algorithm based on direction energy distribution. The saving of computation time for the first rough mode decision can be up to 68% under the negligible PSNR drop and bit rate increase.

I. INTRODUCTION With the presence of HD and Ultra HD video content, Joint

Collaborative Team on Video Coding (JCT-VC) develops a new video coding standard called High Efficiency Video Coding (HEVC) [1]. HEVC provides 33.7% better coding efficiency than the state-of-the-art video codec H.264/AVC by adopting a lots of new features [2]. One of these new features is increasing the number of intra prediction modes up to 35 including Planar mode, DC mode, and 33 angular modes as shown in Fig. 1 (a). However, to find the optimal mode, exhaustive search on these 35 prediction modes leads to much more computation time.

Several fast intra mode decision algorithms are proposed to reduce the computation time of the encoder. The Piao’s algorithm adopted in HM1.0 which is the HEVC reference software first determines the N best candidate modes by testing all 35 modes with sum of absolute Hadamard transformed difference (SATD) [3]. Then the rate-distortion optimization (RDO) on the N candidate modes are performed to find the best intra mode. In addition to choosing N candidate modes, Zhao’s algorithm adopted in HM2.0 adds the most probable mode (MPM) selected from the neighboring prediction units (PU) to improve the coding quality [4]. The Yan’s algorithm reduces the number of candidate modes from 3 to 1 for the PU size bigger than 8x8 by exploring the correlation of the selected candidates in the first rough decision [5]. Although the number of candidate modes for RDO process are reduced, it still requires much time to test all 35 modes during the first rough decision.

Silva exploits edge information of the image to speed up the first rough decision [6]. It first calculates one non-directional edge strength and four directional edge strength: horizontal, vertical, 450, and 1350 on a 4x4 PU. Then only the strongest edge orientation and its fixed eight adjacent directions are chosen for the first rough decision.

This paper presents a fast intra mode decision algorithm based on direction energy distribution. First, the four main

direction energy at degree 0, 45, 90, and 135 are calculated. Then construct a set according to the correlation of energy strength in the four directions. Only 11 or 13 candidate modes rather than 35 are selected for the first rough decision dependent on the direction distribution in the set. The experimental result shows that up to 68% intra mode computation complexity is reduced with 0.04db BD-PSNR degradation and 0.65% BD-rate increase.

(a) (b) Fig. 1. (a) Intra prediction modes in HEVC (b) Intra mode decision algorithm in HM6.1

The rest of this paper is organized as follows. In Section 2, we present the proposed fast intra mode decision algorithm. Experimental results are shown in Section 3. Finally, this paper is concluded in Section 4.

II. PROPOSED FAST INTRA MODE DECISION ALGORITHM Fig. 1 (b) describes the intra mode decision algorithm in

HM6.1. To avoid exhaustively checking all intra modes in the first rough decision, this paper proposes a coarse-to-fine intra mode decision algorithm. In the first stage, it performs the coarse mode decision based on the direction energy distribution described below:

Step 1. Calculate four main direction energy at degree 0, 45, 90 and 135 like [7]. An NxN-size PU is equally divided into four N/2xN/2-size parts and an N/2xN/2-size central part. The energy of a part is calculated by accumulating N/2*N/2 pixels. �� � � � ���

� �

���

� �

��� , ���� � � � ���

� � ���

� ���

�� � � � ���� ��

��

��� � � �� � � � ���� ���

� ��� ,

�� � � � �����

��

����

��

��

���� , ���� ��� !�"#!$ �#%��&'�%�&(�)�� *+ in a NxN PU

1 Fast Intra Mode Decision for HEVC Based on Direction Energy Distribution

Ce-Min Fang, Yuan-Teng Chang, Wen-Hao Chung Industrial Technology Research Institute, Taiwan R.O.C.

A B

C DF

N

N

978-1-4673-6199-6/13/$31.00 ©2013 IEEE

2013 IEEE 17th International Symposium on Consumer Electronics (ISCE)

61

��, ��� -�� . ��- / -�� . ��-������0�, ��� -�� . ��- / -�� . ��- ��1, � -�� . ��- / -�� . ��-������ 21, � -�� . ��- / -�� . ��- Step 2. Choose the direction with minimum energy (Emin) as the best direction. Then put the best direction and the directions with similar energy to it into a set A. By our experiment, it is suitable to set the threshold 3 as 10. �4�5 � 6�(7�88� �9:8�� �;88�� �<=:8> ?@A�B � 7C�-�-�� . �4�5- D �3� � E 78�� 9:�� ;8�� <=:�>�> Step 3. Select 7 candidates from 33 angular modes based on the direction distribution in set A. If there is only one direction in set A, it represents that the strongest direction is found and unique. In this case, this angular mode as well as the six adjacent angular modes with an interval of two modes are chosen as shown in Fig. 2 (a). Otherwise, it means no obvious or unique strongest direction. In this case, seven or nine candidate modes will be selected according to the cardinality of set A discussed below: Case 1. |A| = 4 It means that the difference of energy strength in any two of these four directions are quite small; that is, no specific direction in this PU can be assured. Therefore, the nine fixed modes frequently used in general image content are chosen as shown in Fig. 2 (b). Case 2. |A| = 2 or 3 If similar energy strength is in two directions, the central direction between them will be chosen as initial mode. Expanding from the initial mode in clockwise and anti-clockwise direction, one mode is chosen in every two modes as shown in Fig. 2 (c). Otherwise, if three directions have similar energy strength, four central mode among them will be chosen in addition to the three directions as shown in Fig. 2 (d). In this case, seven candidate modes are adopted.

Fig. 2. Examples for different choices of candidate modes based on set A (a) there is only one best direction at 900 (b) |A|=4 (c) A={900, 1350} (d) A={00, 900, 1350} (e) selection of second initial modes Then it calculates the SATD cost of the seven or nine candidate modes. The mode with minimum SATD is chosen

as the second initial mode. In the second stage, the SATD cost of Planer mode, DC mode, and the two additional modes adjacent to the second initial mode as shown in Fig. 2 (e) are calculated. Finally, depending on PU size three or eight small SATD modes from the above results are selected for RDO process like HM6.1.

III. EXPERIMENTAL RESULT The algorithm presented in this paper has been realized in

HM6.1. Instead of testing all 35 modes, it just selects 9 or 11 modes based on edge energy distribution in four main directions. The encoder is tested by 14 HD video sequences from [8] set as all-intra main configuration. The coding performance is evaluated in terms of BD-rate and BD-PSNR under QP 20, 24, 28, and 32. The simulation result in Table I shows that it just leads to less than 1.39% BD-rate increase and 0.11db BD-PSNR degradation.

TABLE I COMPARISON OF CODING PERFORMANCE WITH HM6.1

Sequence ΔPSNR ΔBitrate Sequence ΔPSNR ΔBitrate ChinaSpeed -0.07 0.76% BQTerrace -0.02 0.45% SlideEditing -0.11 0.77% Cactus -0.02 0.49% SlideShow -0.11 1.20% Kimono -0.01 0.29% vidyo1 -0.06 1.39% ParkScene -0.01 0.20% vidyo3 -0.03 0.73% Tennis -0.01 0.35% vidyo4 -0.03 0.70% Traffic -0.03 0.51% BasketballDrive -0.01 0.50% PeopleOnStreet -0.04 0.73%

IV. CONCLUSION This paper proposes a new fast intra mode decision

algorithm for HEVC encoding. This algorithm only chooses 9 or 11 candidate modes from 33 angular modes by analyzing the energy distribution in four main directions. Finally, it reduces 68% first rough mode computation complexity with negligible PSNR drop and bit rate increase.

REFERENCES [1] B. Bross, W. J. Han, J. R. Ohm, G. J. Sullivan, Y. K. Wang, and T.

Wiegand, “High Efficiency Video Coding (HEVC) text specification draft 10,” JCTVC-L1003, Geneva, Jan. 2013

[2] B. Li, G.J. Sullivan, and J. Xu, “Comparison of Compression Performance of HEVC Draft 8 with AVC High Profile and Performance of HM8.0 with Different Delay Characteristics,” JCTVC-K0279, Shanghai, Oct. 2012

[3] Y. Piao, J.-H. Min, and J. Chen, “Encoder improvement of unified intra prediction,” JCTVC-C207, Guangzhou, Oct. 2010

[4] L. Zhao, L. Zhang, S. Ma, and D. Zhao, ”Fast mode decision algorithm for intra prediction in HEVC,” in Proc. VCIP, 2011

[5] S. Yan, L. Hong, W. He, and Q. Wang, "Group-Based Fast Mode Decision Algorithm for Intra Prediction in HEVC," in Proc. STITS, 2012

[6] T. L. Silva, L. V. Agostini, and L. D. S. Cruz, "Fast HEVC intra prediction mode decision based on edge direction information," in Proc. EUSIPCO, 2012

[7] A. C. Tsai, J. F. Wang, W. G. Lin, and J. F. Yang, "A Simple and Robust Direction Detection Algorithm for Fast H.264 Intra Prediction," in Proc. ICME, 2007

[8] F. Bossen,“Common test conditions and software reference configurations,” JCTVC-F900, Torino, Jul. 2011

2013 IEEE 17th International Symposium on Consumer Electronics (ISCE)

62