Motion-compensated temporal filtering with optimized motion estimation

The lifting-based discrete wavelet transforms applied temporally along the motion trajectory is a powerful tool in scalable video compression. The longer filter is used for motion-compensated temporal filtering (MCTF) for better de-correlation. The bidirectional motion estimation (ME) used for MCTF is the major speed bottleneck. This paper propose two stages motion predictive technique to optimize the ME. In the first stage, zero motion block prediction is carried out using the content of high frequency frame, obtained after TF without motion compensation. In the second stage, the inter-frame and inter-layer prediction technique is applied to skip the coarse search of any fast ME, except for the first pair of frames in a group of pictures. The proposed algorithm is implemented in MC-EZBC. The variations observed in objective or subjective quality of the decoded video is negligible compared to the speed gained.