Fuzzy-Based Motion Vector Smoothing for Motion Compensated Frame Interpolation

Abstract

SUMMARY We proposed a new motion vector (MV) smoothing usingfuzzy weighting and vector median filtering for frame rate up-conversion.A fuzzy reasoning system adjusts the weighting values based on the localcharacteristics of MV field including block difference and block boundarydistortion. The fuzzy weighting removes the affect of outliers and irregularMVs from the MV smoothing process. The simulation results show that theproposed algorithm can efficiently correct wrong MVs and thus improvevisual quality of the interpolated frames better than conventional methods. key words: fuzzy reasoning, motion vector smoothing, motion compen-sated frame interpolation, motion estimation 1. Introduction In the video processing, motion information described bymotion vectors (MVs) is one of the most important factorsfor many video coding applications such as object tracking,video concealment, or motion compensated frame interpo-lation. Unfortunately, tracking real motion is an ill-posedproblem. The motion estimation may be failed wheneverthere are multiple local minima in the sum of absolute dif-ferent distribution due to deformable motion, noise, objectocclusion, lighting variation, and so on. As a result, the in-terpolated frame using wrong motion vectors suffers blockartifact or ghost effect.Many motion vector processing schemes have beenproposed to derive true motion from the estimated MVs.It is possible to use the assumption that the MV fields aresmooth function. The simplest method for MV smoothingis vector median filter [1] which can simply remove MV out-liers and refine MVs from their neighborhood. Alparone etal. proposed an adaptively weighted vector median filter us-ing the measure of displaced frame difference (DFD) to ob-tain smoother MVs with true spatial correlation of the MVfield [2]. Dane et al. presented a method adopting medianand low pass filtering with MV classification based on anglevariance [3]. In the work [4], Ai-Mei Huang et al. proposedbidirectional MV processing by selecting the best MV basedon the minimum difference between forward and backwardpredictions. All these methods try to evaluate the reliabilityof MVs before applying smoothing function. However, therelation between the reliability of MVs and frame difference