1. Ai-Mei Huang and Truong Nguyen Video Processing LabECE Dept, UCSD, La Jolla, CA 92093 This paper appears in: Image Processing, 2007. ICIP 2007. IEEE International Conference on

2. Overview  Introduction  Block diagram of the proposed algorithms  Prediction residual energy analysis  The proposed multi-stage motion vector processing method  Simulation results  Conclusions

3. Introduction  Motion-compensated frame interpolation (MCFI) improves temporal quality by increasing the frame rate at the decoder.  Frame interpolation for compressed video remains a problem due to the use of improper MVs are often generated.  The proposed algorithms preserve the object structure information but also produce a smoother motion vector field (MVF).

4. Block diagram of the proposed algorithms

5. Prediction residual energy analysis(1/5)  In [4], we have discussed that there exists a strong correlation between MV reliability and its associated residual energy.  These high residual energies regions are distributed over where object edges are located.  Let v m,n denote the MV of each 8×8 block. We classify v m,n into three different reliability levels, reliable, possibly reliable.

6. Prediction residual energy analysis(2/5)  For a MB (16×16) with only one MV, we simply assign the same MV to all four 8×8 blocks(b m,n ):  If E m,n ≧ ε 1, it will be considered as unreliable(L1).  Consider intra-coded MBs as unreliable(L1).  The neighboring of L1 MBs or MVs in the same MB will be classified as possibly reliable (L2).  Other MBs will be classified as reliable (L3).

7. Prediction residual energy analysis(3/5) Motion Vector Reliability Map

8. Prediction residual energy analysis(4/5)  Analyze the connectivity of the unreliable MVs in MVRM and create a MB merging map.  If a MB that has unreliable MVs connecting to other unreliable MVs in vertical, horizontal or diagonal directions in adjacent MBs, these MBs will be merged.  The merging process is performed on a MB basis using MVRM, and all MBs will be examined in a raster scan order.  The 32×32 block size is the maximum for merging.

9. Prediction residual energy analysis(5/5)  The diagonal direction is not considered for intra- intra MB merging, because the possibility for two diagonal intra-coded MBs belonging to the same object is lower. (c) MV reliability classification map. Unreliable and reliable MVs are marked in yellow and white colors, respectively. Intra-coded MBs are marked in cyan color. (d) MB merging map.

10. Block diagram of the proposed algorithms

11. The proposed multi-stage motion vector processing method(1/4)  Find the best MV for each merged group:  If the ABPD of v* b is less than a threshold ε 2, assign v* b to the merged MBs in Cu.  Otherwise, drop the selected MV(v* b ) and wait until a proper MV propagates to its neighborhood in next iteration.  Process stops until all merged groups have been assigned new MVs. S denotes the reliable MVs in merge group & adjacent blocks. C u denotes the merged group.

12. The proposed multi-stage motion vector processing method(2/4)  Reclassify MV reliability based on BPD resulted from the selected MV.  BPD(m, n) of each 8×8 block is obtained by simply summing up difference error like Eq(1).  If BPD(m, n) is higher than ε 3,v m,n is unreliable(L1).  Otherwise, other MVs will be classified as reliable(L2).  If the MB consists of multiple motion, those unreliable MV can be easily detected by BPD.

13. Block diagram of the proposed algorithms

14. The proposed multi-stage motion vector processing method(3/4)  For those unreliable MVs of 8×8 blocks in the updated MVRM, correct them by using a reliability and similarity constrained vector median filter: S contains the neighboring MVs centered at v m,n d i,j denotes the distance between v i,j and v m,n V m,n

15. The proposed multi-stage motion vector processing method(3/4)  For those unreliable MVs of 8×8 blocks in the updated MVRM, correct them by using a reliability and similarity constrained vector median filter:  Two MVs are considered to be similar if the angle distance(d i,j ) is below a threshold, ε4.  Before updating v* m,n in MVF 2, check on BPD of v ∗ m,n to ensure that error energy is descended.  If the energy check fail, correct it in next iteration. S contains the neighboring MVs centered at v m,n d i,j denotes the distance between v i,j and v m,n

16. Block diagram of the proposed algorithms

17.  MV smoothing process in [3] to reduce visual artifacts due to high BPD.  On the frame boundary, using unidirectional interpolation based on the directions of MVs: The proposed multi-stage motion vector processing method(4/4)

18. Simulation results(1/3)

19. Simulation results(2/3)

21. CONCLUSIONS  We propose a novel algorithm based on the received information for MCFI.  Accomplishing the concept of object motion without complex motion estimation.  The method outperforms other conventional methods on both objective and subjective video quality.