Computation-Scalable Multi-Path Motion Estimation Algorithm

Transcription

1 JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 9, (0) Computation-Scalable Multi-Path Motion Estimation Algorithm KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI Department of Electrical Engineering National Dong Hwa University Hualien, 97 Taiwan Department of Video Coding Core Technology Industrial Technology Research Institute Hsinchu, 0 Taiwan Controlling the computational complexity to balance video quality and computation power consumption in motion estimation is a significant issue. This paper proposes a fast multi-path motion estimation algorithm for achieving computational complexity reduction as well as controllable computation. First, an enhanced interested diamond search pattern with search point reduction is proposed. In addition, the multi-path search concept is further combined in our proposal to search for the best matched block without suffering from the drawback of being trapped at the local minimum. Finally, by combining the adaptive threshold for early termination, the computation and coding performance can be adaptively controlled by our proposed algorithm to satisfy the various needs of video encoders. Experimental results show that the proposed algorithm with the ability of computation control can give better performance than the diamond search and its multi-path combined algorithm. Keywords: motion estimation, video coding, block-matching, diamond search, intersected diamond search, multi-path search, early termination, computation scalability. INTRODUCTION Motion estimation, a widely adopted technique in many video coding standards, has been utilized to efficiently reduce the temporal redundancy in video data [-]. In motion estimation, the current encoding block has been employed to search for the best matching block in the previous reference frame by an exhaustive search process called full search (FS) motion estimation. Once the best matching block in the previous reference frame has been found, only the displacement between the current encoding block and the previous reference block accomplished with block intensity differences is encoded. Thus, video compression is achieved. However, although the motion estimation algorithm can significantly eliminate temporal redundancy, it places a remarkable computational demand on FS. Therefore, investigating efficient and low complexity motion estimation algorithms has garnered lots of attention in the video coding research field. Much of the literature has proposed reducing the computational requirements of motion estimation [-8]. Although the computational burden can be reduced efficiently, it still remains high in terms of motion estimation. In addition, these literatures only focus on the computational complexity reduction without considering the constraints of computation power availability. As a result, it is not applicable for battery equipped Received January 0, 0; accepted April 9, 0. Communicated by Ruay-Shiung Chang and Sheng-Lung Peng. 079

2 080 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI mobile devices. [9-] discuss the computation scalability for motion estimation and the performance of motion estimation can be varied according to the computational controlling factor. Developing an efficient computation scalable motion estimation algorithm is still a hot topic in the research field. For the latest video coding standard, high efficiency video coding (HEVC) [] targeting HDTV application, the problem of high computational complexity motion estimation further results in noticeable computation power consumption. In addition, for the power limited mobile device, the high computational load of motion estimation leads to a shorter operating life of mobile devices. Therefore, if the computational load of motion estimation can be further reduced, not only will the operating life of mobile devices be extended, but also the goal of green computing can be achieved. This paper proposes an enhanced version of an intersected diamond search incorporating a multi-path and computation scalable concept for motion estimation. In addition, a criterion for early termination of the search of motion estimation is further combined in our proposed algorithm to reduce the computational requirement. The rest of this paper is organized as follows. In section, the algorithms of a new intersected diamond search (NIDS) and multi-path motion estimation are briefly described because these algorithms are the fundamentals of our proposal. Our proposed algorithm is introduced in detail in section. Section shows extensive simulation results to demontrate the efficiency of our proposed algorithm. The conclusion is given in section 5.. OVERVIEW OF NEW INTERSECTED DIAMOND SEARCH AND MULTI-PATH SEARCH. New Intersected Diamond Search (NIDS) Fig. shows the search patterns used in NIDS [5] in which the ISP and SDSP are the abbreviations of the intersected search pattern and small diamond search pattern, respectively. The operation of a NIDS is briefly described as follows. (a) ISP (b) SDSP Fig.. Illustration of the search patterns of the new intersected diamond search (NIDS). Step : The ISP is applied to evaluate the sum of absolute difference (SAD) of each candidate position. Go to Step. Step : If the candidate position with the smallest SAD is located at the center, go to Step. Otherwise, go to Step. Step : The SDSP is applied to compute the SADs for another four candidate positions

3 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 08 and the candidate position with the smallest SAD will be selected as the best result. The search process is terminated. Step : If the candidate position with the smallest SAD doesn t occur at the center position in the ISP step, the pair-wised direction determination will be used to decide which direction the following search operation will move to (as shown in Fig. ). Here, the pair-wised direction determination is according to the minimum of the average SAD of each position pair. Go to Step 5. Step 5: Once the moving direction of search process has been decided, set up the new search center and go to Step. (a) (b) (c) (d) Fig.. Pair-wised direction determination of NIDS.. Multi-Path Search To avoid the search process being trapped into local optimal, the multi-path search approach [6] provides a way which allows the search process to check multiple paths to find the best results. Fig. shows an example to illustrate how the multi-path search is applied to the diamond search (DS) [9] pattern. For a multi-path search, more than one of the candidates with the least SADs will be selected as the search centers after evaluating an initial search pattern and the search will thus be spread out Fig.. Illustration of multi-path search [6] applied to DS.

4 08 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI. PROPOSED ALGORITHM The proposed algorithm is mainly composed of three components; () Enhanced Intersected Diamond Search (EIDS); () multi-path EIDS, and; () computation scalable search and dynamic early termination approach. The details of each component are introduced as follows.. Proposed EIDS By reducing search points in search pattern of DS, NIDS requires less computation than DS and achieves competent performance. Nevertheless, from the operation of NIDS, we can observe that the ISP is applied to the entire search process until the minimum SAD position occurs at the center. As described in [], most real-world sequences have a centrally biased motion vector distribution, which means that the best motion vector could be easily found around the center [7, 8]. However, the NIDS ignores many checking positions during its search operation as represented by cross symbols in Fig.. As a result, it might be possible that the best result would be skipped over and thus degrade the prediction accuracy. Our proposed EIDS is a modified search pattern which tries to resolve the potential drawback of the NIDS algorithm Fig.. Illustration of skipped candidates of NIDS. Fig. 5 shows a step-by-step example to illustrate our proposed EIDS algorithm. In the first step, the operation of our proposal is similar to the NIDS algorithm (Step ()). Once the pair-wised direction determination process has decided the search direction, the position (labeled by square S in gray in Step ()) that has not been checked by the NIDS algorithm will be evaluated by our EIDS algorithm. Afterwards, the SADs of the filled circle labeled as and S are compared to determine which one has the minimum SAD value. If S is not the position with minimum SAD, the NIDS algorithm will be applied to the rest of the search processes (Step (a) to Step (5a)) until the best result has been found. However, if S has the minimum SAD, the SADs of neighboring positions around S are compared to determine which position has the minimum SAD among them. If the

5 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 08 upper position has minimum SAD (as shown in Step (b)), this position will be treated as the new search center and the two additional positions(labeled as in Step (b)) will be checked. From now on, our enhanced pattern with S checking point will be applied in the rest of the search processes to find out the best result. The benefit of our proposed EIDS is that our proposal not only takes the low computation complexity advantage from NIDS but also avoids the miss-checked positions nearby, which potentially could be the best result. S Step () Step () SAD Min at S S S Step (a) Step (a) Step (5a) SAD Min at S S S S S S Step (b) Step (b) Step (5b) Fig. 5. Illustration of the proposed EIDS algorithm.. Proposed Multi-Path EIDS Algorithm To further increase the coding performance, we incorporate the multi-path search strategy into our proposed EIDS algorithm. Fig. 6 shows an example to illustrate how the multi-path concept works when combined with our proposed EIDS algorithm. The priority of search centers for the paths is according to the order of minimum SADs of the search points shown in Step in Fig. 5.

6 08 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI (a) -path (b) -path (c) -path (d) -path Fig. 6. Illustration of our proposed multi-path EIDS. Fig. 7 exhibits the flowchart of the proposed multi-path EIDS algorithm. The detailed coding flow is described as follows. First, some variables are initialized for the following motion estimation usage. The initialized variables are introduced as follows. NP: the number of supported search paths p: the index for indicating the current processing path SAD Min [p]: the variable for storing the minimum SAD of path p MV Best [p]: the variable for storing the best MV of path p After the initialization process, the ISP is applied to the first step to derive the initialized search results. If the minimum SAD is located at the center position after the ISP search, the search pattern is changed to SDSP and the best motion vector (MV) is thus obtained from the SDSP search results. The SDSP consists of five checking points and forms with small diamond shape as in Fig. (b). Otherwise, our proposed multi-path EIDS algorithm will be executed. For each search path, the EIDS will be conducted to find the best motion vector and SAD. Once all of the best MVs and SADs for all paths have been successfully discovered, the MV Best will be selected from one of all searched paths with the smallest SAD value.

7 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 085 Start of motion estimation Initialization for variables NP = # of paths; p = ; SAD Min = SAD Min [i] = ; MV Best [i] = (0,0); <= i <= NP; Execute the first step of EIDS SAD Min located at center Next SP for the pth path of EIDS Derive SAD Temp and MV Temp SAD Temp < SAD Min [p] SAD Min [p] = SAD Temp MV Best [p] = MV Temp Last SP of EIDS of path p p = p + p <= NP Find SAD Min from SAD Min [i], <= i <= NP Derive MV Best from MV Best [i] Execute SDSP Derive MV Best End of motion estimation operation Fig. 7. Flowchart of the proposed MPS-EIDS algorithm.

8 086 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI. Proposed Multi-Path EIDS Algorithm with Computation Scalable and Early Termination Due to the diversity of end user computing capabilities, the motion estimation algorithm with constant computational requirement is no longer capable to meet the demands of computing power variation. Even worse, motion estimation algorithms without the consideration of current computing power capability might sometimes lead to the crash of the video coding system because the motion estimation module may run out of overall computing power before it finishes the video coding process. Therefore, there is a great demand to design an efficient motion estimation algorithm which also takes into consideration the current computing power capability to dynamically adjust its motion estimation computational complexity. Fig. 8 shows the flowchart of the proposed multi-path EIDS algorithm with computation that is scalable. In our proposal, the SP Used and the computational controlling factors SP Control are denoted to indicate the used number of search points so far and the available number of search points, respectively. Compared to Fig. 7, we can observe that the main difference between these two figures is the insertion of the determination of computational controlling factors after every search point checking of SDSP and EIDS. During the execution for both the multi-path EIDS and SDSP, if the SP Used has arrived at the computational controlling factor SP Control, it means that the current computing capability is insufficient and the search process would be terminated immediately. Otherwise, the search process is continued for the next search point until the termination condition has been met. In our proposal, the computational controlling factor SP Control is either user controllable or dynamic system operating power probing depending on the practical application. Fig. 9 shows the search order of EIDS pattern given the constraint of search points. To further increase the motion estimation speed, an early termination mechanism [5] has been adopted in our proposal. From Fig. 8, we can find that every SAD Temp of every search step of every path has to be determined to see whether the search for the current path has to be terminated or not. If SAD Temp is less than the dynamic threshold, Th, the rest of the search for that path will be skipped. In our proposal, the variable of threshold, Th, is derived by the following formula. Th = Mean(MinSAD Left, MinSAD Top, MinSAD TopRight ) β () where MinSAD Left, MinSAD Top, and MinSAD TopRight are the minimum SAD values of left, top and top right MB, respectively. The parameter β is adjustable according to the available computation ability. Once all of the best results of all paths have been found out, the best motion vector will be selected as the MV Best.. EXPERIMENTAL RESULTS This section exhibits several simulation results to demonstrate the coding efficiency of our proposed algorithm. Full search (FS), new intersected diamond search (NIDS), diamond search (DS), multi-path diamond search (MPS-DS) and our proposed algorithm

9 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 087 Start of motion estimation Initialization for variables NP = # of paths; p = ; SAD Min [i] = ; SP Used = 0; MV Best [i] = (0, 0); <= i <= NP; Calculate Th; Execute the first step of EIDS SAD Min located at center Next SP for the pth path of EIDS Derive SAD Temp and MV Temp Execute next SP for SDSP SP Used = SP Used + SAD Temp < SAD Min [p] SP Used < SP Control SAD Min [p] = SAD Temp MV Best [p] = MV Temp SP Used = SP Used + Last SP of SDSP w/o ET SP Used < SP Control w ET Derive MV Best SAD Temp <Th Last SP of EIDS of path p p = p + p <= NP Find SAD Min from SAD Min [i], <= i <= NP Derive MV best from MV Best[i] End of motion estimation Fig. 8. Flowchart of proposed computation scalable motion estimation algorithm.

10 088 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI Step (5a) Step (5b) Fig. 9. Illustration of the search order for the computation scalable EIDS according to Fig. 5. are included for comparison. Tables - show the experimental results for several test sequences which contain SIF Football, CIF Table Tennis (TT), CIF Soccer, 70p Pedestrian and Rush-hour. In these tables, the term of ProxP stands for our proposed algorithm incorporating x-path search. For example, ProP means that two paths are adopted in our proposed EIDS algorithm. SP Control stands for the controlling factor of the number of search points enclosed by brackets. In addition, the term ET means that the early termination mechanism is applied. In our simulation, the search range is set to ±7 and 0 frames are used for encoding. Each macroblock (MB) consists of 6 6 pixels and gets one MV for each algorithm. Table. PSNR(dB) comparisons for different algorithms. SIF CIF CIF 70P Football TT Soccer Pedestrian Rush-hour Avg. FS DS NIDS MPS-DS ProP ProP ProP ProP ProP+SP Control (0) ProP+SP Control (5) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (5)+ET(β=0.) ProP+SP Control (5)+ET(β=0.5) ProP+SP Control (5)+ET(β=.0) ProP+SP Control (5)+ET(β=.5) ProP+SP Control (5)+ET(β=.0) Average PSNR and the number of search point comparisons are shown in Table and Table, respectively. Our proposed algorithm(prop) without incorporating a multipath search and early termination can individually achieve.0db and.89db PSNR

11 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 089 Table. Average search points per MV comparisons for different algorithms. SIF CIF CIF 70P Football TT Soccer Pedestrian Rush-hour Avg. FS DS NIDS MPS-DS ProP ProP ProP ProP ProP+SP Control (0) ProP+SP Control (5) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (5)+ET(β=0.) ProP+SP Control (5)+ET(β=0.5) ProP+SP Control (5)+ET(β=.0) ProP+SP Control (5)+ET(β=.5) ProP+SP Control (5)+ET(β=.0) increases on average compared to DS and NIDS algorithms. For our proposed algorithm (ProP) combined with -path search,.78db and.7db PSNR improvements can be achieved compared to DS and NIDS, respectively. When the early termination algorithm has been adopted in the proposed algorithm to further increase motion estimation speed, our proposal can still maintain PSNR performance. On average, 0.8dB and.07db PSNR gain can be acquired by our proposed algorithm(prop+sp Control (5)+ET(β=.0)) compared to DS and NIDS, respectively. Regarding the MPS-DS algorithm, since this algorithm searches multi-paths to find the best result, a higher PSNR performance can be obtained but at the expense of high computational complexity as shown in Table. It can be seen that the average number of search points per MV of MPS-DS algorithm is higher than that of our proposed algorithm. We can observe that our proposed algorithm can provide different combinations of PSNR and computational requirements by adjusting β and SP Control. For the β value, a higher β means a higher error tolerance and also means a lower prediction quality but less search points. ProP+SP Control (5)+ET(β=.0) can aim at a much lower computational complexity requirement with acceptable PSNR degradation. However, for high quality application, we can use ProP+SP Control (5)+ET(β=0.) to derive higher PSNR results. By properly selecting β and SP Control values, our proposed algorithm can achieve quality as well as computational complexity scalability. Compare the experimental results of ProP+SP Control (0) and ProP+SP Control (5)+ (β=0.), the latter is tend to get better PSNR than the former because of the early termination mechanism. If the early termination criterion during the search process has accomplished, the search process can switch to the next path until all the paths have been searched or the SP Control has been run out. However, the former has to try all search points in current path and then continue to the next path until all the paths have been searched or

12 090 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI the SP Control has been run out. More paths can be searched for the latter if the early termination criterion is easy to reach, thus the PSNR of motion estimation could be better for the latter one which has more opportunities to get the best search result. Table. MSE comparisons for different algorithms. SIF CIF CIF 70P Football TT Soccer Pedestrian Rush-hour Avg. FS DS NIDS MPS-DS ProP ProP ProP ProP ProP+SP Control (0) ProP+SP Control (5) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (0) ProP+SP Control (5)+ET(β=0.) ProP+SP Control (5)+ET(β=0.5) ProP+SP Control (5)+ET(β=.0) ProP+SP Control (5)+ET(β=.5) ProP+SP Control (5)+ET(β=.0) Table exhibits the MSE comparisons for different algorithms. It can be seen that our proposed algorithm without incorporating multi-path and early termination mechanisms can obtain less MSE than those of DS and NIDS algorithms. In our proposed algorithm combined with a multi-path search mechanism, much more MSE improvement can be achieved. Similarly, our proposal can still result in less MSE than DS and NIDS algorithms if the early termination mechanism is adopted. For computational controlled cases ProP+ SP Control (0) and ProP+SP Control (0), the maximum numbers of search points are 0 and 0 for four search paths, respectively. Although each case can search for four paths, the number of search points is actually limited by the SP Control factor. Take the ProP+SP Control (0) case as example; if the first path consumes seven search points, the remaining number of search points is only three. In this situation, only three search points can be used for other paths. However, since the number of search points of ProP+SP Control (0) case is 0, it means that much more search point resources can be dispatched to all search paths. Therefore, the ProP+SP Control (0) can acquire lower MSE when compared to the case of ProP+SP Control (0). Figs. 0- exhibit some performance comparison curves of different algorithms for several sequences. From these figures, we can observe that our proposed algorithm can achieve much better PSNR performance than NIDS and DS algorithms at similar search points. Compared to the MPS-DS algorithm with the highest PSNR performance, our proposed algorithm requires a much lower number of search points and provides almost the

13 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 09 Fig. 0. Performance comparison for various algorithms for SIF football sequence. Fig.. Performance comparison for various algorithms for CIF soccer sequence. Fig.. Performance comparison for various algorithms for 70P pedestrian sequence.

14 09 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI same PSNR performance. For MPS-EIDS, we can find that the PSNR performance improvement between paths and is significant and the greatest performance gain between successive paths is obtained. Therefore, for low computational complexity application, MPS-EIDS with only two paths supplies a good tradeoff between computational complexity and PSNR performance. In addition, we can observe that the NIDS, DS, and MPS-DS can t provide the ability to scale the computational requirement for deriving different encoding quality points. However, our proposed algorithms can provide the computation scalability. For higher quality applications with higher computational demand, our proposed MPS-EIDS algorithm can be used for such applications. However, for a wide range of computational complexity applications, our proposed MPS-EIDS+SP Control +ET would be a good choice. Our proposal can provide a high flexibility for adjusting the PSNR performance and computational complexity when compared to all other algorithms. Although both MPS-EIDS+SP Control +ET and MPS-EIDS+SP Control algorithms utilize SP Control to control the number of allowed search points, the MPS-EIDS+SP Control +ET algorithm has been further combined with an early termination mechanism to stop the search process in the current searching path. When the early termination condition has been met, the search process would be switched to another path for searching until the SP Control has run out or all paths have been searched. However, in MPS-EIDS+SP Control, since no early termination has been included, the SP Control might run out for a single searching path. When both the MPS- IDS+SP Control +ET and MPS-EIDS+SP Control have the same computational resource constraint, most of the PSNR performances of MPS-EIDS+SP Control +ET would be higher than those of MPS-EIDS+SP Control. 5. CONCLUSION In order to compensate for the drawback of the NIDS algorithm, this paper proposes an enhanced intersected diamond search algorithm with step-wise search point reduction for motion estimation. Our proposed algorithm further incorporates the multi-path concept to avoid the motion search process to being trapped into the local optimal. To further achieve computational complexity reduction as well as controllable computation, an early termination technique is also adopted in our proposal. Simulation results show that our proposed algorithm outperforms the well-known DS and NIDS algorithms. Under the similar PSNR condition, our proposed algorithm consumes much less computational complexity than the MPS-DS algorithm. In addition, by changing the search point controlling factor and early termination threshold, our proposed algorithm can provide different PSNR results for various applications. The low computational complexity as well as the computation controllable motion estimation can be provided to support hardware device diversity by proposed algorithm. REFERENCES. T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra, Overview of the H.6/ AVC video coding standard, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 00, pp Advanced video coding for generic audiovisual services, ITU-T Recommendation

15 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 09 H.6, B. Bross, W.-J. Han, J.-R. Ohm, G. J. Sullivan, and T. Wiegand, High efficiency video coding (HEVC) text specification draft 9, Document JCTVC-K00 of Joint Collaborative Team on Video Coding (JCT-VC), October, 0.. C. C. Chang, L. L. Chen, and T. S. Chen, A new two-path search algorithm for block motion estimation of video data, Journal of Information Science and Engineering, Vol. 7, 00, pp K. L. Chung, T. J. Yao, and Y. H. Huang, New prediction- and affine transformation-based three-step search scheme for motion estimation with application, Journal of Information Science and Engineering, Vol., 008, pp R. Li, B. Zeng, and M. L. Liou, A new three-step search algorithm for block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 99, pp M. J. Chen, L. G. Chen, and T. D. Chiueh, One-dimensional full search motion estimation algorithm for video coding, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 99, pp L. M. Po and W. C. Ma, A novel four-step search algorithm for fast block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 6, 996, pp S. Zhu and K. K. Ma, A new diamond search algorithm for fast block-matching motion estimation, IEEE Transactions on Image Processing, Vol. 9, 000, pp Z. Ce, X. Lin, and L. P. Chau, Hexagon-based search pattern for fast block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 00, pp J. Y. Tham, S. Ranganath, M. Ranganath, and A. A. Kassim, A novel unrestricted center-biased diamond search algorithm for block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 8, 998, pp L. K. Liu and B. Feig, A block-based gradient descent search algorithm for block motion estimation in video coding, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 6, 996, pp C. H. Cheung and L. M. Po, A novel cross-diamond search algorithm for fast block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 00, pp X. Jing and L. P. Chau, An efficient three-step search algorithm for block motion estimation, IEEE Transactions Multimedia, Vol. 6, 00, pp Y. J. Lai, C. H. Li, and S. W. Leu, High efficient new intersected diamond search algorithm for fast motion estimation video coding, in Proceedings of International Conference of Digital Technology and Innovation Management, 006, pp S. Goel and M. A. Bayoumi, Multi-path search algorithm for block-based motion estimation, in Proceedings of IEEE International Conference on Image Processing, 006, pp V. Christopoulos and J. Cornelis, A center-biased adaptive search algorithm for block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 0, 000, pp H. Nisar and T. S. Choi, An advanced center biased search algorithm for motion

16 09 KUANG-HAN TAI, GWO-LONG LI, MEI-JUAN CHEN AND HAO-WEN CHI estimation, in Proceedings of IEEE International Conference on Image Processing, Vol., 000, pp H. W. Cheng and L. R. Dung, A power-aware motion estimation architecture using content-based subsampling, Journal of Information Science and Engineering, Vol., 006, pp M. Y. Chiu and W. C. Siu, Computationally-scalable motion estimation algorithm for H.6/AVC video coding, IEEE Transactions on Consumer Electronics, Vol. 56, 00, pp C. S. Kannangara, I. E. Richardson, and A. J. Miller, Computational complexity management of a real-time H.6/AVC encoder, IEEE Transactions on Circuits and System for Video Technology, Vol. 8, 008, pp C. Y. Chen, Y. W. Huang, C. L. Lee, and L. G. Chen, One-pass computation-aware motion estimation with adaptive search strategy, IEEE Transactions on Multimedia, Vol. 8, 006, pp P. L. Tai, S. Y. Huang, C. T. Liu, and J. S. Wang, Computation-aware scheme for software-based block motion estimation, IEEE Transactions on Circuits and Systems for Video Technology, Vol., 00, pp W. Lin, K. Panusopone, D. M. Baylon, and M. T. Sun, A computation control motion estimation method for complexity-scalable video coding, IEEE Transactions on Circuits and System for Video Technology, Vol. 0, 00, pp F. Y. Chien, An improved fast motion estimation algorithm based on adaptively thresholding and architecture design, Master Thesis, Department of Electrical Control Engineering, National Chiao-Tung University, Taiwan, 005. Kuang-Han Tai ( ) received the M.S. degree from the Department of Electrical Engineering, National Dong Hwa University, Hualien, Taiwan, in 0. He is now pursuing the Ph.D. degree in the same department. His research interests include video/image compression and motion estimation. Gwo-Long Li ( ) received his B.S. degree from the Department of Computer Science and Information Engineering, Shu-Te University, Kaohsiung, Taiwan, in 00; M.S. degree from the Department of Electrical Engineering, National Dong Hwa University, Hualien, Taiwan, in 006; and Ph.D. degree from the Department of Electronics Engineering, National Chiao- Tung University, Hsinchu, Taiwan, in 0. In 006, he received the Excellent Master Thesis Award from Institute of Information and Computer Machinery. He is currently an Engineer in Indus-

17 COMPUTATION-SCALABLE MULTI-PATH MOTION ESTIMATION ALGORITHM 095 trial Technology Research Institute (ITRI), Hsinchu, Taiwan. His research interests are the video signal processing and its VLSI architecture design. Mei-Juan Chen ( ) was born in Taiwan in 969. She received the B.S., M.S., and Ph.D. degrees in Electrical Engineering from National Taiwan University, Taipei, in 99, 99, and 997, respectively. She was an assistant professor ( ) and an Associate Professor ( ) in the Department of Electrical Engineering, National Dong Hwa University, Hualien, Taiwan. Since August 005, she has been a Professor of the Department of Electrical Engineering, National Dong Hwa University. She also served as the Chair of her department from Her research topics include image/video processing, video compression, motion estimation, error concealment and video transcoding. In 99 and 997, she was the recipient of the Dragon Paper Awards and the Xeror Paper Award. She was also the recipient of 005 K. T. Li Young Researcher Award from ACM Taipei/Taiwan Chapter for her contribution on video signal codec technique. This award is given annually to only one person under the age of 6, conducting research in Taiwan. In 006, she received the Distinguished Young Engineer Award from the Chinese Institute of Electrical Engineering, Taiwan. In 006, she also received the Excellent Master Thesis Supervision Award from Institute of Information and Computer Machinery. In 005 and 0, she received Jun S. Huang Memorial Foundation best paper awards. She serves as associate editor for EURASIP Journal on Advances in Signal Processing since 008. Hao-Wen Chi ( ) was born in Taiwan in 989. He received the B.S. in Electrical Engineering from Tamkang University, Taipei, Taiwan, in 0 and received the M.S. degree from the Department of Electrical Engineering, National Dong Hwa University, Hualien, Taiwan, in 0. His research topics include image/video processing, video compression, motion estimation and multi-view video coding.