New block motion estimation algorithm for video compression C.-K. Ning, K.-T. Lo and H.-T. Tang Centre for Multimedia Signal Processing Department of Electronic and Information Engineering The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong enktlo@polyu.edu.hk Abstract: Most of the existing fast block matching algorithms like three-step search (TSS) and fourstep search (4SS) are optimized only for smaller search window. Their MSE performance will be degraded when the window size is getting larger, which is unreasonable that a poorer performance is obtained when more efforts are done. To tackle this problem, a new quadrant based four-step search (QB4SS) algorithm is developed in this paper. This algorithm refines the searching steps of the 4SS so as to adjust the searching pattern to suit for different local characteristics of the block. Experimental results show that the proposed algorithm achieves much better MSE performance than other search methods when the search range is set to ± 15. Keywords: Video compression, motion compensation, motion estimation [2-10]. However, most of these fast search algorithms like the three-step search (TSS) do not encounter the local characteristic of the test block, and only a regular search pattern is applied for all kinds of motion. Moreover, they preassume the motion of the scene is mainly centerbiased in [6]. Actually, most of the motions in real life may not be center-biased. On the other hand, the search pattern of such algorithms is optimized for small search window only. Their MSE performance will be degraded when the search window becomes larger. In this paper, a quadrant based four-step search (QB4S) algorithm is proposed for block matching motion estimation. The QB4S uses different search patterns for blocks with different characteristics. For each testing block, the motion type is first determined and the corresponding search pattern is applied. With this arrangement, the probability of trapped by local minima will be reduced. Introduction Motion Compensation is one of the efficient interframe coding techniques for its ability to reduce the high redundancy between successive frames of an image sequence. Motion compensation consists of two main processes: motion estimation and prediction error coding. Motion estimation is to obtain the movement between successive frames and the resulting motion information is then exploited in interframe predictive coding. Both the motion information and the prediction error are requested to transmit to the receiver for image reconstruction. The better and higher speed of the motion estimation process is, the more efficient the encoding process is. In the last two decades, various motion estimation techniques have been proposed. Among them, the most popular method so far has been the block matching motion estimation technique. This approach is used in various coding standards like MPEG [1]. The block matching (BM) method is to find out a candidate block, within a search window in the reference frame, that is most similar to the test block in the current frame. The full search method tests all locations in the search window and produces the optimal results. If the performance in terms of prediction error is only the criterion for BMA, the full search method is optimal. However, its large computational requirement makes it difficult to be realized in hardware or software. Therefore, investigation of fast search algorithms has been the focus of research in the last decade In the rest of the paper, section 2 describes the concept of the proposed fast algorithm and the details of the quadrant based four-step search (QB4SS) algorithm are discussed in section 3. Computer simulations are used to evaluate the performance of the new algorithm and the results are presented in section 4. Finally, some concluding remarks are given in section 5. Concept of the Quadrant Based Four-Step Search Algorithm In general, most of the fast search algorithms [210] are based on two very important assumptions. The first assumption is that MAD will increase 84 monotonically as the checking point moves away from the global minimum and the second is the principle of locality. With the multiple stage search schemes like TSS, the checking points in each step are allocated uniformly in the search window. Such a configuration may not be appropriate for those blocks having small motions. Also, most of the fast algorithms, their probability of trapped by local minima will increase when search range increases. The reasons for that phenomenon are: 1) The reliability of monotonic assumption and principle of locality for the error surface plane are greatly reduced; and 2) Continuing reduce in step size will greatly accumulate the error especially for large motion. Step 3: If the minimum point is located at the middle of horizontal or vertical axis of the pervious search windows, use the searching pattern of four-step search to locate the motion vection. The step is demonstrated in Fig.1b. Step 4: If the minimum point is located at the corner of horizontal or vertical axis of the pervious search windows, most likely the block will have a large motion and the motion vector will be located in the quadrant of the minimum point. In this case, a hierarchical search is designed to locate the motion vector in this quadrant. For the hierarchical search, the search points in the target quadrant will be sub-sampled by a factor of 2 (or 4 when the search range is doubled). All marked points are then checked. When the minimum point is found, eight more neighboring points are checked to conclude the search and locate the motion vector. The step is illustrated in Fig.1c. On the other hand, some search schemes such as new three step search (NTSS), eight extra points are added in the first step to determine whether the motion of the block is center-biased or not. Such configuration is not good for large motion. Therefore, how to optimally design a universal search pattern becomes the main concern of this paper. In our algorithm, the first task is to determine whether the motion of the block is center-biased or not. This can be accomplished by checking the location of the point having smallest mean absolute difference (MAD) value in the nine test points at the first step. If the minimum point is at the center point, then most likely the motion will be center-biased. It should be a stationary and quasi-stationary motion block. Otherwise, it belongs to fast motion block. In that case, we need to determine it belongs to diagonal, horizontal or vertical motion. Quadrant Based 4-Step Search In the proposed QB4S algorithm, a center-biased search pattern with 9 checking points on a 55 window is utilized in the first step. Then the search method of next step is depended on the location of the minimum points. The QB4S algorithm is summarized as follows: Figure 1 Different search paths of QB4S Simulation Results Step 1: Search the 9 checking points on a 55 window Computer simulations have been performed to compare the performance of the proposed Quadrant Based 4-step step (QB4S) with full search (FS), three step search (TSS), new three step search (N3SS), four step search (4SS), orthogonal search (OSA) and 2D_logrithms Search (2D-Log). Two test sequences, namely “Miss America” and “Tennis”, are used in our Step 2: Find out the point with minimumt MAD value. If the location of the minimum point is in center, use the searching pattern of the block based gradient descent search (BBGDS) to locate the motion vector, which is illustrated in Fig.1a 85 experiments. They are of CIF standard and the simulations have been performed on the first 90 frames of the sequences. In our experiment, the maximum motion displacement W is set to ± 7 and ± 15. The performance of the motionpredicted images is measured by mean square error (MSE), and the computational complexity is evaluated by the average number of search point per frame. Tennis with 15 search window 800 FS 700 MSE per pixel TSS 600 N3SS 500 4SS 400 Ortho 300 2D_log 200 QB4S 85 79 73 67 61 55 49 43 37 31 25 19 7 1 0 13 100 frame number Figure 2 shows the MSE performance of different search methods for the two test sequences with different search ranges. Tables 1 to 4 summarize the average MSE performance and the average number of search points for various search methods for the two test sequences. (d) Tennis with search range ± 15 Figure.2 MSE performance of different search methods Missa with 7 search window MSE per pixel 16 FS TSS N3SS 4SS Ortho 2D_log Q4BS FS 14 TSS 12 NtSS 10 4SS 8 Ortho 6 2D_log 4 QB4S 2 Average Search points/frame 225 25 21.67 19.249 13.0 15.105 18.966 Average MSE per frame 8.177295 8.83 8.2779 8.6937 9.818 8.67 8.5856 85 78 71 64 57 50 43 36 29 22 8 15 1 0 Table 1 Average MSE performance for Miss America with search range ±7 FrameNumber (a) Miss America with search range ± 7 FS Missa with 15 search window N3SS FS TSS N3SS 4SS Ortho 2D_log Q4BS 4SS Ortho 2D_log 89 81 73 65 57 49 41 33 25 17 9 QB4S 1 MSE per pixel TSS 16 14 12 10 8 6 4 2 0 Average Search points/frame 961 33 21.61 27.56 17.0 19.27 22.8030 Average MSE per frame 8.10 9.025 8.295 8.79 9.889 8.83 8.5819 frame number Table 2 Average MSE performance for Miss America with search range ±15 (b) Miss America with search range ± 15 Tenni s wi t h 7 search wi ndow FS 600 TSS N3SS 400 4SS 300 Ort ho 200 2D_l og 100 QB4S 89 81 73 65 57 49 41 33 25 9 17 0 1 MSE per pixel 500 f rame number FS TSS N3SS 4SS Ortho 2D_log Q4BS Average Search points/frame 225 25 22.62 20.106 13.00 16.105 19.175 Average MSE per frame 180.5639 239.00 215.56 210.4068 296.79 215.42 206.85 Table 3 Average MSE performance for Tennis with search range ±7 (c) Tennis with search range ± 7 86 Acknowledgement FS TSS N3SS 4SS Ortho 2D_log Q4BS Average Search points/frame 961 33 23.15 28.69 17.0 20.154 32.852 Average MSE per frame This work was supported by The Hong Kong Polytechnic University under Grant A/C G-V466 and the Centre for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University. 158.57 258.632 236.62 229.09 333.63 236.69 196.227 References [1] ISO/IEC 11172-2 (MPEG-1 Video), “Information Technology - Coding of Moving Pictures and Associates Audio for Digital Storage Media at up to about 1.5 Mbit/s: Video”, 1993. [2] T. Koga, K. Linuma, A. Hirano, Y. Iijima, and T. Ishigr, “Motion compensated interframe image coding for video conferencing,” Proc. NTC81, pp. G5.3.1-5.3.5, New Orleans, LA, Nov. 1981 [3] J. R. Jain and A. K. Jain, “Displacement measurement and its application in interframe image coding”, IEEE Trans. Commun., vol. COM-29, pp. 1799-1808, Dec 1981. [4] S. Kappagantula and K. R. Rao, “Motion compen-sated interframe image prediction,” IEEE Trans. Commun., vol. COM-33, pp. 10111015, July 1985. [5] M. Ghanbari, “The cross-search algorithm for mot-ion estimation”, IEEE Trans. Commun., vol.38, no. 7, pp. 950-953, July 1990. [6] R. Li, B. Zeng, and M. L. Liou, “ A new three step search algorithm for block motion estimation," IEEE Trans Circuits and Systems for Video Tech., vol.4, no.4, pp.438-442, Aug. 1994. [7] L.W. Lee, J. F. Wang, J. Y. Lee, and J. D. Shie, “Dynamic search-window adjustment and interlaced search for block-matching algorithm," IEEE Trans. Circuits and Systems for Video Tech., vol.3, no. 1, pp. 85-85, Feb. 1993. [8] L. M. Po and W. C. Ma, “A novel four-step search algorithm for fast block motion estimation," IEEE Trans. Circuits and Systems for Video Tech., vol. 6, no. 3, pp. 313-317, June 1996. [9] B.Zeng, R. Li and M.L. Liou, “Optimization of fast block motion estimation algorithms," IEEE Trans. Circuits and Systems for Video Tech., vol. 7, no.6, pp 833-844, Dec. 1997. [10]K. L. Liu and E. Feig, “A block-based gradient descent search algorithm for block motion estimation algorithms in video voding,” IEEE Trans. Circuits and Systems for Video Tech., vol. 6, no.4, pp 419-422, Dec. 1996. Table 4 Average MSE performance for Tennis with search range ±15 It is seen that the FS method achieves the lowest MSE, and our proposed QB4S method performs a very little worse than FS. When compared with other methods, our proposed QB4S method performs better than TSS, OSA, 2D-Log and 4SS in all cases since it requires less number of search points but also results in smaller MSE than those methods. It is also noted from the tables that the performance all search algorithms degrades when the search range is increased from ±7 to ±15. The degradation will become significant when the sequence is of large motion like “Tennis”. It is consistent with the fact that most of the search algorithms are based on the center-biased motion assumption. For our proposed QB4SS algorithm, the MSE performance is improved when the search window increases. Take the sequence “Tennis” as an example, the average MSE of the predicted images is 206.85 and 196.227 search range of ±7 and ±15 respectively. So we may conclude our proposed algorithm is robust and can handle sequences with different motions. Conclusions In this paper, a new search algorithm called quadrant based four-step search (QB4SS) algorithm is developed for block matching motion estimation. The search pattern of the QB4SS algorithm is adjusted according to the local characteristics of the block and is not optimized only for small search window. Experimental results show that QB4SS not only achieves compatible MSE performance with other existing algorithms, but also perform much better than others when the search window is getting larger. The improvement is more significant especially for the sequences with large motions. 87