Implementation of fast motion estimation algorithms and comparison with full search method in H.264

Transcription

1 IJCSNS International Journal of Coputer Science and Network Security, VOL.8 No.3, March Ipleentation of fast otion estiation algoriths and coparison with full search ethod in H.264 A.Ahadi, M.M.Azadfar Departent of Planning & Engineering, Mobile Counications Copany of Iran (MCCI), Tehran, Iran Multiedia Systes Group, IT Faculty, Iran Telecounication Research Center (ITRC), Tehran, Iran Suary In this paper, Three otion estiation algoriths for H.264 standard have been ipleented and perforance and soe features of these three have been tested. These algoriths are full search and two fast search ethod. Finally soe paraeters such as bit rate, speed up and PSNR for different value of QP have been copared. Ipleentation has been perfored in Matlab software. Key words: ITU-H.264 standard,otion estiation,full search,fast search. 1. Introduction Motion Estiation (ME) is an iportant part of any video copression syste, since it can achieve significant copression by exploiting the teporal redundancy existing in a video sequence. Unfortunately it is also the ost coputationally intensive function of the entire encoding process. In otion estiation the current iage is divided into Macro Blocks (MB) [1]. Most of algoriths have been proposed for otion estiation use fro BMA_based (Block Matching Algoriths) ethods. In this ethods, otion estiation is perfored for a N M blocks of current frae, It is done with checking entire N M blocks fro search area situated in the reference frae(s) and calculating the difference between the current block and other reference blocks and finally choosing the block that has the ost siilarity (iniu SSD) to the early block in current frae. Then, difference of two blocks as residual (otion copensated residual) and the distance of the as otion vector, is coded and transitted. Search area in the reference frae, generally is a range of [- 15,15] [-15,,15] pixels relative to the location of the block to be searched. SSD A = F (, ) ( s) F G sεa in SSD G 2 ( s) (1) In H.264, unlike other standard, M and N is variable and are selected based on coplexity value of otion. The seven possible odes of a block in H.264, are 16 16, 16 8, 8 16, 8 8, 8 4, 4 8, 4 4. Fitting selection of the block sizes leads to acceptable quality in low bit rates. The proble with ethod of choosing the best ode of block size is the tie, because of the process coplexity, we cannot find the best in real tie. The otion estiation process accounts for ore than 50% of total encoding tie in case of one reference frae, however, as the nuber of reference fraes increases, the relative coputational portion of otion copensation increases gradually, and at last it is about 70% in case of four reference pictures.[2] In fast search algoriths, the otion estiation process instead of full search, follows special pattern that checks less point nuber, Such as diaond pattern (figure 1) and hexagonal[1] pattern. saller otion copensation block sizes can produce better otion copensation results. However, a saller block size leads to increased coplexity (ore search operations ust be carried out) and an increase in the nuber of otion vectors that need to be transitted. Sending each otion vector requires bits to be sent and the extra overhead for vectors ay outweigh the benefit of reduced residual energy. An effective coproise is to adapt the block size to the picture characteristics, for exaple choosing a large block size in flat, hoogeneous regions of a frae and choosing a sall block size around areas of high detail and coplex otion.[3] The ost iportant criterion for this selection is ratedistortion (RD) and iniization the j value that achieves fro equation (2). j = SSD + λ r (2) Where SSD is su of square error and λ is equal to: ( QP 12) / 3 λ = (3)

2 140 IJCSNS International Journal of Coputer Science and Network Security, VOL.8 No.3, March 2008 Value of QP is between 0 and 51 and r represents the nuber of bits that needs for otion vector coding. The inforative RD optiization technology provides a bitrate reduction of up to approxiately 10% and a PSNR iproveent of up to 0.35 db, while it increases the coputational coplexity of the H.264 encoder.[4] n 2 (2 1) PSNRdB = 10log10 (4) MSE checking points of LDSP are tested at the first step. If the iniu SAD occurs at the search center of LDSP, then the search pattern is switched fro LDSP to SDSP, and the position having the iniu SAD in SDSP is decided as the otion vector. Otherwise, a new center of LDSP is placed at the point that yields the iniu SAD in the previous step, and all points on the new LDSP are tested again. This process is iteratively repeated until the iniu SAD falls on the search center. 2. MB ode deterination and otion estiation ethods The seven possible odes of a block, as entioned earlier, are 16 16, 16 8, 8 16, 8 8, 8 4, 4 8, 4 4. Fitting selection of the block sizes leads to acceptable quality in low bit rates. Very different ethods ay be used in this section that soe of the in the following 2.1. Full Search ethod (recoendation of standard) Generally, in this ethod, all possible odes is checked. With perforing the otion estiation for every blocks and calculating the R-D criterion for all of the, We can decided which block sizes should be used. First of all, otion estiation for acroblock (16 16 block) is perfored and j is calculated. Then acroblock is divided into two 16 8 and then 8 16 blocks and for each of the, otion estiation and j calculation is perfored. The su of calculated j of blocks in each ode, is the j of that ode. Fro this four state, a ode that has iniized the j is choosed. If the selected ode is 8 8, breaking process of each blocks is continued like the previous. The sallest possible block sizes is 4 4 and afterward the breaking procedure is finished. The ost iportant disadvantage of this algorith versus it s high precision is, the over tie expended to procedure and this,overshadow the feature of realtie coding. 2.2.The first ethod of fast search and MB ode deterination (fast search 1) [2] In this algorith, fast otion estiation with odified diaond search for variable block sizes is perfored. Motion vector field adaptive search technique (MVFAST) uses a different initial search point and search patterns with selective application of large diaond search (LDS) and sall diaond search (SSD) according to the characteristics of otion activity assessed by the siilarity of otion vector field aong contiguous blocks. In the LDS case, the LDSP (LDS pattern) shown in Fig.1 is centered at the search center, and all the (a) LDSP (b) SDSP Fig. 1 Two types of diaond search pattern In the proposed schee, fro the siilarity of otion vector aong hierarchical block size structure is utilized instead of neighboring block s otion field as in MVFAST for the otion field adaptive search. ME is perfored fro 4 4 to block sizes. At first, sixteen basic otion vectors of all 4 4 blocks in one acroblock are found by MVFAST ethod. The search centers and patterns for larger blocks of 8 4, 4 8, and 8 8 are selected on the basis of the otion vector deviation of the 4 4 blocks constituting the given block. The otion vector deviation D of a given block of 8 4, 4 8, 8 8 is coputed as the average absolute distance between 4 4 block otion vector and their ean vector as: N 1 D = MVb( i) ean _ vector (5) N i= 1 Where, N is the nuber of basic otion vectors of MVb(i) in the given block, and MVb(i) is the otion vector of i th 4 4 block in a given block. If all basic otion vectors of 4 4 block corresponding to the current block are the sae, that is, if D equals to zero, the ean vector becoes the otion vector of current block, and ME is skipped. If D is saller than or equal to T 1, only the first step of the sall diaond search (SDS) is perfored at the center position. That is, only five points as shown in Fig. 1(b) is tested. Otherwise, if D is saller than or equal to T 2, the SDS is perfored. If D is larger than T 2, LDS is perfored. Therefore, as otion vector deviation increases, ore coplex and larger diaond search starting fro the

3 IJCSNS International Journal of Coputer Science and Network Security, VOL.8 No.3, March ean vector is applied. The sae process is also perfored for 16 8, 8 16, blocks with regard to 8 8 block otion vectors. In this experient, the values 1 and 8 are used respectively for the thresholds T 1 and T 2. [2] 2.3. The second ethod of fast search and MB ode deterination (fast search 2)[5] This algorith has been proposed to reduce the nuber of potential odes and to restrict the set of past coded reference picture for ME. The algorith will eliinate ME for soe block types and reference pictures and will differentiate Skip ode (SKIP ode refers to the ode where no otion and residual inforation is encoded, So no otion search is required and it has the lowest coplexity) fro other block types and give it the highest priority. The algorith is based on whether the error surface versus block size is onotonic, that is, whether the current acroblock has the sae tendency of using saller block size (sub-acroblock partition) or larger block size. The error surface is built by initial 3 odes (block sizes): 16 16, 8 8, and 4 4. Here 8 8 eans that the entire acroblock is exained using only 8 8 partitions, and 4 4 eans that the entire acroblock is exained using only 4 4 partitions. We call that the error surface is onotonic if : J ode (16_16)< J ode (8_8) <J ode (4_4) or versa. If the error surface is not onotonic, all other odes need to be tested. If the error surface is onotonic, only odes (block sizes) between the best two odes are tested. For exaple, if the best two odes are and 8 8, which iplies that the acroblock tends to use larger block partitions, only 16 8 and 8 16 are further tested; if the best two odes are 8 8 and 4 4, this iplies that the acroblock tends to use saller block partitions (or sub-acroblock partitions), and only 8 4 and 4 8 odes are further tested. step1: check SKIP ode. if J ode (SKIP) < T1, select SKIP as best ode, stop; otherwise go to step2; step2: check and 8 8. if (J ode (SKIP) < J ode (16_16))&&(J ode (SKIP)< J ode (8_8)), go to step7; otherwise, go to step3; step3: check 4 4; if ((MinJ ode = = J ode (8_8)) (MaxJ ode = = J ode (8_8))), go to step4; if MaxJ ode = = J ode (4_4), go to step5; if MaxJ ode = = J ode (16_16), go to step6; step4: check 16 8, 8 16, 8 4, 4 8; go to step7; step5: check 16 8, 8 16; go to step7; step6: check sub-acroblock partition; go to step7; step7: choose the best ode aong all tested odes. In step1, threshold T1 = N bits.λode, where N bits equals the iniu nuber of bits required for non SKIP inter odes. In step2, by coparing SKIP with and 8 8, we assue that if the RD cost for SKIP is the inial, then the probability for other odes to have cost less than SKIP will be very sall, so no other odes need to be checked. We check the onotonic condition in step3, In step6, an additional decision is perfored for each 8 8 partition to decide which type shall be used aong the 4 sub-acroblock partitions. Only 8 4 and 4 8 need to be tested. The search for otion estiation in this algorith also is perfored according to LDS pattern. 3. Ipleentation results The experients were carried out on a 1.6 MHz Intel Pentiu achine with 256 MB RAM eory. The environent for coparing the perforances is MATLAB software. The algoriths is applied on two successive fras of the bitstrea. The first frae as reference frae is used for coding the next frae. For exaple, figure 2 shows two successive fraes of silent bitstrea and difference of the(residual) has been shown in figure 3. In the residual frae, PSNR is db and j is equal to but with applying full search algorith on the fraes, otion copensated residual frae shown better results. PSNR and J are db and 1950 respectively. This process last about 54 seconds in test states. This frae is shown in figure 4. After running The fast search(1) algorith on the fraes, following result (fig 5) will be achieved. In this frae, PSNR is db and average of J criterion is 2561, while the Speed_up value is 4.9 respect to the full search ethod. In the next state, with applying the fast search(2) algorith, PSNR, J and Speed_up will be db, 4590 and 7.7 respectively. Residual frae has been shown in figure 6. The energy and bit_rate of this frae is larger than previous and these are the cost of better Speed_up. All of these tests is perfored for QP=26. Calculation of speed_up paraeter is as following: Tie Full search _ (6) TieFast search Speed up = Fig. 2 two successive fraes of silent bitstrea

4 142 IJCSNS International Journal of Coputer Science and Network Security, VOL.8 No.3, March 2008 These successive two fraes have been chosen fro forean, coastguard, hall, news and silent bitstreas and their paraeters such as bit-rate, speed_up and PSNR have been copared. Both of bit-rate and speed_up paraeters is respect to the Full search and hence one value has been assigned to the. The bit-rate calculation is perfored based on RD cost value in the picture. The coparison is presented for three separate values of QP paraeter (consist of 5, 26, 50). the results is shown in annex tables(fro table1 to table4). To obtain ore precision results Table4 for 10 successive fraes (F=10)(at QP=26) has been achieved. Fig. 3 difference of two fraes Table 1: coparison at QP=5 (F=1) Table 2: coparison at QP=26 (F=1) Fig. 4 residual frae after applying full search Table 3: coparison at QP=50 (F=1) Table 4: coparison at QP=26 (F=10) Fig. 5 residual frae after applying fastl search(1) Fig. 6 residual frae after applying fastl search(2) Fig. 7 forean bitstrea

5 IJCSNS International Journal of Coputer Science and Network Security, VOL.8 No.3, March Fig. 8 Coastguard bitstrea Fig. 9 news bitstrea Fig. 10 hall bitstrea 4. Conclusion In this paper, three different algoriths for otion estiation are tested and copared. The fraes are I picture type and otion estiation process is perfored using one reference frae.coparison between FS(1) and FS(2) : Whereas FS(1) operates with sall block sizes, In sall value of QP and fraes with a onotonic otion (forean), it is faster than FS(2) and produces lower bit-rate. But when the otion details or oving areas decrease, FS(1) has lower speed up than FS(2), because FS(2) first checks larger blocks and has the skip ode. For larger QP or less otility fraes, it has faster operation and rather than others shows better bit-rate. When QP increases, The speed up value for FS(1) is decreased, but in FS(2), this is increased strongly. And it is due to the starting search procedure fro sall block sizes in FS(1) and fro larger block sizes in FS(2). And for all QP values, Full search algorith has the best PSNR, FS(2) has the ost speed up and PSNR for FS(1) is better than FS(2). As it sees, when QP increases, the PSNR is reduced for all algoriths, because this leads to choosing larger block sizes and produces a significant aount of energy in otion copensation residual frae. For sall QPs, The ost aount of bit-rate is for FS(2) and for large QP the ost is for FS(1). References [1] M. Ghanbari, Standard Codecs : Iage Copression to Advanced Video Coding. IEE, [2] Paolo De Pascalis, Luca Pezzoni, Gian Antonio Mian, and Daniele Bagni, "fast otion estiation with size-based predictors selection hexagonal search in H.264/AVC encoding", IEEE 2003 [3] Woong IL Choi; Byeungwoo Jeon; "fast otion estiation with odified diaond search for variable otion block sizes", International Conference on Iage Processing, Volue 2, pp 371-4, Sept [4] H.264 and MPEG-4 Video Copression, Iain E. G. Richardson, John Wiley Press, [5] Ki-Hun HAN, Yung-Lyul, "Fast Macroblock Deterination to Reduce H.264 Coplexity", IEICE TRANS. Fundaentals, Vol.E88 A, No.3 MARCH 2005 [6] Peng Yin; Tourapis, H.-Y.C.; Tourapis, A.M.; Boyce, J.; "Fast ode decision and otion estiation for JVT/H.264", International Conference on Iage Processing, Volue 3, pp 853-6, Sept [7] T. Shanableh and M. Ghanbari, Heterogeneous video transcoding tolower spatio-teporal resolutions and different encoding forats, IEEETrans. on Multiedia, vol. 2, no. 2, pp , Jun [8] R. Srinivasan and K. R. Rao, Predictive coding based on efficientotion estiation, IEEE Trans. on Co., vol. 2, no. 2, pp ,1985. [9] H. G. Musann, P. Pirsch, and H. J. Grallert, Advances in picture coding, Proc. IEEE, vol. 73, pp , Apr [10] M. Kung, VLSI Array Processors. Englewood Cliffs NJ: Prentice Hall,1987. [11] K. Tavassoli and W. Badawy, A prototype for parallel otion estiation architecture using full-search block atching algorith, in International Workshop on Digital and Coputational Video (DCV 02), Nov [12] J.-C. Tuan and C.-W. Chang, T.-S. Jen, On the data reuse and eory bandwidth analysis for full-search blockatching vlsi architecture, IEEE Trans. on Cicuits and Systes for Video Tech.,vol. 12,No.1,pp.61 72,Jan Air Ahadi received the B.E. degree in Counications, fro Tabriz Univ. (Tabriz-IRAN) in He also received his M.S degree in Electronics fro Iran University of Sience & Technology (Tehran-IRAN) in He has worked as a Planner in Dept. of Planning & Engineering at Mobile Counications Copany of Iran (MCCI) since His research interest includes Digital Iage Processing, Iage and Video Coding. M.Mahdi Azadfar received the B.E. degree in Electronics fro Ferdowsi Univ. (Mashad-IRAN) in He also received his M.S degree in Electronics fro Tarbiat Modaress Univ.(Tehran-IRAN) in He has worked as a researcher in Dept. of Multiplex,Network, Inforation Technology, at Iran Teleco Research Center (ITRC) since His research interest includes Digital Iage Processing, Iage and Video Coding, Analog and Digital Television, Multiedia systes, Video Broadcasting over Network. He is currently with the ITRC as a faculty eber.