Fast Intra Prediction for High Efficiency Video Coding Hao Zhang1 and Zhan Ma2 1

School of Information Science and Engineering Central South University, Changsha, Hunan 410083 China Email: [email protected] 2 Dallas Technology Lab Samsung Electronics, Richardson, TX 75206 USA Email: [email protected]

Abstract. Emerging High Efficiency Video Coding (HEVC) video coding standard promises the significant compression performance improvement compared to the H.264/AVC. However it comes with the tremendous encoding complexity increase. Thus, it is very useful and necessary to develop fast algorithms for HEVC, so as to reduce the encoder complexity. In this paper, we propose a fast intra prediction scheme for HEVC to reduce the prediction mode search for each prediction unit. For all available test sequences provided by the JCT-VC, it demonstrates 38% encoding time reduction for all intra case with BD-RATE increase about 2.9%. Several sub-algorithms are developed and integrated for complexity reduction. First, we applied the Hadamard transform on 2:1 downsampled prediction residual to derive the sum of absolute Hadamard transformed difference (SATD) for rough mode decision, where a progressive search process is then used to reduce effective mode candidates for fully rate-distortion optimized quantization (RDOQ). Finally, an early termination based on SATD cost and mode distances is also included in RDOQ process to further complexity reduction. Extensive simulation results demonstrate that our proposed method is quite efficient for intra mode prediction speed-up. Our proposal is complementary to other separated works on fast coding unit, prediction unit, and transform unit decision. We expect more encoder complexity reduction by combing our solution and other fast algorithms.

1

Introduction

The emerging video coding standard HEVC [1], under the efforts of Joint Collaborative Team on Video Coding (JCT-VC), has achieved significant compression efficiency in comparison to the widely deployed H.264/AVC [2]. Although it is still under the block-based hybrid motion-compensation and transform coding framework, HEVC is much more complex by introducing the recursive tree structure, large block transforms, advanced motion prediction, additional filtering operations and etc [1, 3, 4]. The macroblock concept has been extended by defining Coding Unit (CU), Prediction Unit (PU) and Transform Unit (TU). Starting

2

from the largest CU (referred as LCU), each CU can be recursively split into multiple sub-CUs. Each leaf sub-CU can be further split into multiple PUs. For the intra case, both CU and PU are squared size with either one 2Nx2N block or 4 NxN sub-block after further splitting, N ∈ [4, 8, 16, 32]. Besides, recursive TU is also implemented on top of each PU for residual coding, which incurs more complexity requirement. Figure 1 gives the simple illustration for recursive CU, PU and TU adopted in HEVC. Please refer to [1] to more detailed description.

Fig. 1: Recursive block structure for HEVC, where k indicates the depth for CUk and TUk .

In addition to the complexity increase introduced by the recursive block structure of HEVC, the complexity is further augmented for intra prediction by using extended spatial prediction directions. As of the Draft 6 [1], there are 35 directional intra prediction candidates, where mode 0 is Planar mode, and mode 1 is DC mode, as shown in Figure 2. It requires unbearable computing resource to conduct full Rate-Distortion Optimized Quantization (RDOQ) for all these modes at different PU levels. In practise, a three-step mode decision process is adopted in the HEVC reference software HM [5]. First, a rough mode decision is performed to select a mode candidate set based on prediction residual SATD and estimated mode bits3 . Reduced number of modes selected in the first stage are chosen for the normal computational intensive RDOQ to obtain the best prediction. Furthermore, recursive transforms are applied to the optimal mode from the first two stages for the residual coding to derive the final best coding mode. With such three-step encoding process optimization, overall intra coding complexity is partly reduced without noticeable quality loss. However, the complexity is still very high due to extensive modes search. In this paper, we propose a novel scheme for fast intra prediction in HEVC, which includes multiple-stage optimization (i.e., Hadamard transform on 2:1 3

In HM, 8x8 and 4x4 Hadamard transforms are applied.

3

Fig. 2: Intra Prediction Modes

downsampled prediction residual, progressive mode search and early RDOQ termination), yielding significant intra coding time reduction at a sacrifice of negligible bit rate increase. The remainder of this paper is organized as follows. Section 2 gives a brief literature review of fast intra prediction methods developed for HEVC. Section 3 describes the proposed fast intra prediction mechanism. Section4 presents experimental results to demonstrate the effectiveness of the proposed solution. Finally, conclusion remarks and discussion are drawn in Section 5.

2

Fast Intra Mode Decision for HEVC: A Review

Fast intra mode decision (FIMD) has been extensively studied for H.264/AVC [6–17]. However, these algorithms can not be directly applied to HEVC due to quite different coding structures and prediction modes. Recently, a few fast intra mode decision (MD) algorithms are proposed for HEVC (some algorithms are proposed for both inter and intra MD). These works are briefed as follows. In HEVC, residual quad-tree coding (RQT) is applied to further exploit the spatial correlation and improve the coding efficiency [1] by implementing recursive TU splitting. However, it demands significant computational overhead. Hence, Tan et al. proposed fast RQT algorithms for both intra and inter mode coding so as to reduce the complexity [18]. As reported using HM2.0, for all intra case, fast RQT saves 13% encoding time with 0.1% BD-Rate increase. For random access and low delay scenarios, proposed fast RQT algorithm reduces up to 9% encoding time at the expense of (up to) 0.3% BD-Rate performance degradation. Another RQT coding scheme is proposed by Teng et al. that replaces the

4

original depth-first mode decision process by a Merge-and-Split decision process [19]. They propose to early terminate merge and split processes, i.e., when the current TU is a zero-block, no further split is performed any more. It reports almost 2x speedup for random access using High Efficiency configuration (for Class C and Class D non-HD sequences) on HM2.0, with negligible coding loss. Meanwhile, Choi et al. have developed a tree pruning algorithm for fast coding unit decision, based on the observation that if the current CU chooses the SKIP mode as the best mode, then no further splitting is required [20]. This algorithm is implemented and reported on HM3.0 software, and claims 40% encoding time reduction. However, early SKIP mode decision does not apply to all intra case. Zhao et al. studied the impact of the number of mode candidates after the rough mode decision [21]. Basically, proposed method tried to reduce the number of candidates for full RDOQ process. According to the experimental data, with a negligible loss of compression efficiency, it achieved 20% and 28% encoding time saving on average in High Efficiency and Low Complexity test conditions using HM1.0 reference software, respectively. A gradient based FIMD is proposed by Jiang et al., where gradient directions are calculated and histogram is generated, for CU size decision [22]. With this approach, 20% time savings on average is achieved with negligible loss of coding efficiency on HM4.0. Shen et al. recently proposed a fast mode decision scheme based on Bayesian decision rule [23]. The split and non-split decision is made on the Bayesian risk, which can be calculated from the Lagrangian cost, the class-conditional probability density functions and priori probabilities. The feature vectors are calculated online while other parameters are calculated off-line. Random Access and Low Delay Configuration are used for simulations and on average 41.4% encoding time saving is reported. Tian and Goto proposed a content adaptive FIMD based on texture complexity analysis and two mode filtering stages [24]. In the pre-stage, LCUs are firstly downsampled to 16x16 blocks, and texture complexity are calculated. The intuition is that when the complexity is less than some threshold, small PUs are not checked; otherwise large PUs are not checked. There are some other mechanisms to filter out small PUs when encoding the last 32x32, e.g., if the minimum PU sizes of its neighboring 32x32 CUs are the same (MxM), then the PUs sizes smaller than M will be skipped. The proposed algorithm achieves averagely 44.91% for 4kx2k sequences and 28.8% for 1080p sequences. Most of the aforementioned algorithms are mainly focusing on early CU, or PU block size decision without requiring further split or merge process, while our scheme is developed to reduce the number of intra prediction modes at any PU level. These algorithms are complementary to our proposal, which can be combined with our solution to achieve more speed up.

5

3

The Proposed Fast Intra Prediction Algorithm

Our proposed fast intra prediction method is composed of three major subalgorithms, including Hadamard transform on 2:1 downsampled prediction residual, progressive mode search for rough mode decision and early RDOQ termination. 3.1

Hadamard Transform on Down-sampled Residual

Given a NxN PU, its prediction residual is the difference between original signal and its predictor, noted as r(i, j), i, j ∈ [0, N − 1]. We first apply the 2:1 downsampling filter on residual signal using a very simple average operator as shown in Figure 3. On top of the down-sampled residual signal, we further perform the Hadamard transform to derive the corresponding SATD, where 4x4 and 2x2 block based Hadamard transforms are used instead of original 8x8 and 4x4 based transforms in HM. As we can see, for each available intra prediction modes, we will have different prediction residual as well as its SATD together with estimated mode bits consumption. 1 2 3 4

(1 +

(a) 8x8 residual block

2

+

3

+

4

) >> 2

(b) average operator

(c) 4x4 residual block

Fig. 3: Illustration of simple averaging based downsampling on a 8x8 prediction residual block, other NxN blocks have the similar operation.

3.2

Progressive Mode Search

Besides using Hadamard transform on down-sampled prediction residual, we also propose the progressive mode search for rough mode decision. Such idea is enlightened by the fast motion estimation (FME) process. For FME algorithms, such as three-step search or diamond search, locations with a large search range are examined first [25, 26]. When the best location at one step is found, it becomes the new search center for the next step. Usually the search range gets smaller at later steps and converges to the best location. First, we define the difference between intra mode indices to represent distance between two modes. For instance, we call mode mi is a d-distance neighbor of mode mj if |mi − mj | = d. During the rough mode decision process, it checks equally spaced eight modes (i.e., two adjacent modes are of distance 4). For

6

the best four modes (with least four SATD cost), check their adjacent modes that are of distance 2 with them. Last, check the adjacent modes of the best two modes that are of distance 1 with them. Each step, we maintain the ordered SATD costs list and their corresponding intra prediction modes, noted as CSATD (m) with m for the intra mode index . It is very like the FME where the search range gradually decreased. 3.3

Early RDOQ Termination

After performing the rough mode decision, reduced number of prediction candidates, i.e., M modes with least costs, are categorized together as Ψ and go through the fully RDOQ process to decide the final best mode mopt . Here, we propose an early RDOQ termination to further encoder time reduction. For each intra mode m ∈ Ψ , we derive its overall cost J(m) as the combination of CSATD (m) and associated mode index bits consumption. Within Ψ , we can have the minimal Jmin for a certain intra mode. Such mode is defined as rough best mode mopt rough . If such rough best mode is Planar or DC mode, i.e., mopt rough = 0 or 1, all other modes in Ψ will be skipped. Otherwise (i.e., mopt rough 6= 0 or 1 , if |m − mopt rough | > 3, such mode m is skipped also; Meanwhile, if J(m) > αJmin , mode m will not be checked. In this paper, we have α = 1.2. After such early termination procedure, all left modes will be checked by RDOQ. 3.4

Integrated Fast Intra Prediction and Illustrative Example

In this section, we integrate aforementioned sub-algorithms to form the proposed fast intra prediction method, as shown in Figure 4. Additionally, we also gives an illustrative example to clearly demonstrate this procedure. 1) Initially, S1 = {0, 1, 6, 10, 14, 18, 22, 26, 30, 34} are checked. Note that the distance between two modes are 4 except for mode 0 and 1. Suppose the best 6 modes as the result of this process are S2 = {0, 6, 10, 14, 18, 26}. 2) The 2-distance neighbors of mode 6, 10, 14, 18, 26 are {4, 8}, {8, 12},{12, 16},{16, 20},{24, 32}, respectively (we assume mode 0 and 1 do not have neighbors). The collection of 2-distance neighbors of S2 is S3 = {4, 8, 12, 16, 20, 24, 28}. Additionally, suppose the modes of upper and left PUs are S4 = {0, 4}. Check modes in S3 and S4. 3) Suppose the best 2 modes till now after checking S1, S2, S3, S4 are {4, 6}. The 1-distance neighbors of mode 4 and 6 are {3, 5}, {5, 7}, respectively. And hence S5 = {3, 5, 7} are checked. 4) Choose the best M modes as candidates for full RDOQ optimization after checking S1, S2, S3, S4, S5. If M = 3 and the best M modes till now is {4, 5, 6}. Suppose the minimum cost of these 3 modes is J4 , and if HSAD cost of mode 5 is much larger than that of mode 4, i.e., J5 > αJ4 , then mode 5 is not required to perform the full RD optimization. The same principle is followed for checking mode 6.

7

for m = [0, 1, 2+4*i], 1≤i≤8 check CSATD(m); Order CSATD(m) in list Q; end for m in Q6 if m ≠ 0, 1 check CSATD(m) for its 2-distance neighbors; update ordered Q; end end // check upper/left block mode check CSATD(m_upper) and CSATD(m_left); update ordered Q; for m in Q2 if m ≠ 0, 1 check CSATD(m) for its 1-distance neighbors; update ordered Q; end end // select candidates if 8x8, 4x4 prediction unit Select best 3 modes from Q as Ψ else Select best 8 modes from Q as Ψ end

(a) progressive rough mode decision

for m in Ψ // // calculate J(m) = CSATD(m) + λLm; find the m for corresponding minimal J end // mopt_rough Jmin = J(mopt_rough); α= 1.2; if mopt_rough = 0 , 1 skip all other m; else for m in Ψ if (|m - mopt_rough| > 3) && (m ≠ 0, 1) skip mode m; end if J(m) > α Jmin skip mode m end end end // Ψnew // for m in Ψnew Find best mode m through RDOQ end

(b) Early RDOQ Termination

Fig. 4: Proposed Fast Intra Prediction Algorithm

5) Follow the early termination process for the full RDOQ optimization as aforementioned. For those modes m which are not skipped, conventional RDOQ optimization is applied.

4

Experimental Results

This sections presents the experimental results with our proposed fast intra prediction, in comparison to the default HEVC encoding using HM6.0, following the common conditions defined in [27]. All intra encoder setting is simulated to demonstrate proposed algorithm. Class A (4Kx2K), B (1080p), C (WVGA), D (QWVGA) and E (720p) sequences are all used for performance verification. Results are shown in Table 1 with BD-Rate performance. In addition, Figure 5 illustrates four samples with their rate-distortion plots. On average, our proposed solution achieves 38% encoding time reduction for all intra coding with less than 3% BD-Rate increase.

5

Conclusion

We propose a novel fast intra prediction approach in this paper. Compared with the reference software HM6.0, proposed algorithm leads to a great encoding

8

Table 1: Coding efficiency and complexity reduction for proposed fast intra prediction on HM6.0

Class A

Class B

Class C

Class D

Class E

Traffic PeopleOnStreet Nebuta SteamLocomotive Kimono ParkScene Cactus BasketballDrive BQTerrace BasketballDrill BQMall PartyScene RaceHorses BasketballPass BQSquare BlowingBubbles RaceHorses FourPeople Johnny KristenAndSara Overall Enc Time[%]

Luma-Y Chroma-U Chroma-V 2.5% 0.4% 0.4% 2.8% -0.4% -0.2% 0.8% 0.9% 0.6% 0.9% 0.5% 0.4% 1.4% 0.4% 0.4% 1.9% -1.0% -0.7% 2.6% -0.1% 0.1% 2.7% 0.0% 0.2% 2.3% -0.7% -1.3% 3.6% 0.3% 0.5% 3.9% 0.3% 0.4% 3.7% 0.5% 0.5% 2.5% 0.1% 0.1% 4.2% 1.0% 1.2% 4.7% 0.7% 0.8% 4.4% 0.6% 0.4% 3.6% 0.9% 0.9% 3.4% 0.3% 0.3% 3.4% 0.5% 0.4% 3.9% 0.8% 0.9% 2.9% 0.5% 0.5% 62%

Traffic@4K

Kimono@1080p

44 42

42 40

40

38 38 36 34

2

4

6

8

36

10

0.5

1

1.5

2

4

4

x 10

x 10

BQMall@WVGA

RaceHorses@QWVGA

42

PSNR (dB)

40

40

38 36

35 HM6 HM6_FastIntraPred

34 32

0.5

1

1.5 Bit Rate (kbps)

2

2.5

30

1000

2000

3000

4000

4

x 10

Fig. 5: Illustration of Rate-distortion performance for proposed fast intra prediction.

9

speed improvement and negligible bitrate increment and PSNR loss. Our solution is trying to reduce the intra prediction modes at any PU level, therefore, other separated works as reviewed in Section 2 can be combined with our work to further boost the encoding speed. We also would like to open our source code [28] and welcome other researchers to use our code base for their research purpose. As the future study, we will work on inter mode prediction complexity reduction and present the combined fast intra/inter prediction for practical HEVC based video applications.

References 1. B. Bross, W.-J. Han, J.-R. Ohm, G. Sullivan, and T. Wiegand, “High efficiency video coding (hevc) text specification draft 6,” in Doc. JCTVC-H1003 dK. Joint Collaborative Team on Video Coding (JCT-VC) of ITU-T SG16 WP3 and ISO/IEC JTC1/SC29/WG11, November 2011. 2. B. Li and G. J. Sullivan and J. Xu, Comparison of Compression Performance of HEVC Draft 6 with AVC High Profile, JCT-VC Doc. I0409r1, April 27 - May 7 2012. 3. G. Sullivan and J.-R. Ohm, “Recent developments in standardization of high efficiency video coding (hevc),” in Proc. SPIE, vol. 7798, August 2010. 4. D. Marpe, H. Schwarz, S. Bosse, B. Bross, P. Helle, T. Hinz, H. Kirchhoffer, H. Lakshman, T. Nguyen, S. Oudin, M. Siekmann, K. Suhring, M. Winken, and T. Wiegand, “Video compression using nested quadtree structures, leaf merging and improved techniques for motion representation and entropy coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 10, pp. 1676–1687, December 2010. 5. Y. Piao, J. Min, and J. Chen, “Encoder improvement of unified intra prediction,” in JCTVC-C207, October 2010. 6. F. Pan, X. Lin, S. Rahardja, K. P. Lim, Z. G. Li, D. Wu, and S. Wu, “Fast Mode Decision Algorithm for Intraprediction in H.264/AVC Video Coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 15, no. 7, pp. 813–822, July 2005. 7. S.-K. Kwon, A. Punchihewa, D. Bailey, S.-W. Kim, and J. Lee, “Adaptive simplification of prediction modes for h.264 intra-picture coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 58, no. 1, pp. 125–129, March 2012. 8. Y.-H. Huang, T.-S. Ou, and H. Chen, “Fast decision of block size, prediction mode, and intra block for h.264 intra prediction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 8, pp. 1122–1132, August 2010. 9. D. Quan and Y.-S. Ho, “Categorization for fast intra prediction mode decision in h.264/avc,” IEEE Transactions on Consumer Electronics, vol. 56, no. 2, pp. 1049–1056, May 2010. 10. K. Bharanitharan, B.-D. Liu, J.-F. Yang, and W.-C. Tsai, “A low complexity detection of discrete cross differences for fast h.264/avc intra prediction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 10, no. 7, pp. 1250–1260, November 2008. 11. C.-H. Tseng, H.-M. Wang, and J.-F. Yang, “Enhanced intra-4 4 mode decision for h.264/avc coders,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 16, no. 8, pp. 1127–1132, August 2006.

10 12. A.-C. T. et al., “Intensity gradient technique for efficient intra-prediction in h.264/avc,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, no. 5, pp. 694–698, May 2008. 13. H. Zeng, K.-K. Ma, and C. Cai, “Hierarchical intra mode decision for h.264/avc,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 6, pp. 907–912, June 2010. 14. C. Kim, H.-H. Shih, and C.-C. J. Kuo, “Fast h.264 intra-prediction mode selection using joint spatial and transform domain features,” J. Vis. Commun. Image R., vol. 17, pp. 291–310, 2006. 15. J. Lee and H. Park, “A fast mode decision method based on motion cost and intra prediction cost for h.264/avc,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 22, no. 3, pp. 393–402, March 2012. 16. D.-Y. Kim, K.-H. Han, and Y.-L. Lee, “Adaptive single-multiple prediction for h.264/avc intra coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 20, no. 4, pp. 610–615, April 2010. 17. A.-C. Tsai, J.-F. Wang, J.-F. Yang, and W.-G. Lin, “Effective subblock-based and pixel-based fast direction detections for h.264 intra prediction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, no. 7, pp. 975–982, July 2008. 18. Y. Tan, C. Yeo, H. Tan, and Z. Li, “On residual quad-tree coding in hevc,” in IEEE International Workshop on Multimedia Signal Processing (MMSP), October 2011. 19. S.-W. Teng, H.-M. Hang, and Y.-F. Chen, “Fast mode decision algorithm for residual quadtree coding in hevc,” in The Visual Communications and Image Processing (VCIP) Conference, November 2011. 20. K. Choi and E. Jang, “Fast coding unit decision method based on coding tree pruning for high efficiency video coding,” Opt. Eng. 51, 030502, March 2012. 21. L. Zhao, L. Zhang, S. Ma, and D. Zhao, “Fast mode decision algorithm for intra prediction in hevc,” in The Visual Communications and Image Processing (VCIP) Conference, November 2011. 22. W. Jiang, H. Ma, and Y. Chen, “Gradient based fast mode decision algorithm for intra prediction in hevc,” in International Conference on Consumer Electronics, Communications and Networks (CECNet), April 2012. 23. X. Shen, L. Yu, and J. Chen, “Fast coding unit size selection for hevc based on bayesian decision rule,” in Picture Coding Symposium (PCS), May 2012. 24. G. Tian and S. Goto, “Content based hierarchical fast coding unit decision algorithm for hevc,” in Picture Coding Symposium (PCS), May 2012. 25. R. Li, B. Zeng, and M. Liou, “A new three-step search algorithm for block motion estimation,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 4, no. 4, pp. 438–442, August 1994. 26. S. Zhu and K. Ma, “A new diamond search algorithm for fast block-matching motion estimation,” IEEE Transactions On Image Processing, vol. 9, no. 2, pp. 287–290, February 2000. 27. F. Bossen, Common test conditions, JCT-VC Doc. I1100, April 27 - May 7 2012. 28. [Online]. Available: http://vision.poly.edu/∼zma03/opensrc/sourceHM6.zip