Wavelet-Domain Motion Estimation with Initial Search Point Prediction Using Low Bit-Depth Pixels

This paper proposes a fast wavelet-domain motion estimation algorithm with initial search point prediction using pixels of two-bit depth. Firstly, we formalize the procedure of bit-depth transformation by two successive steps, namely interval partitioning and interval mapping. A non-uniform quantization method is employed to compute three coarse thresholds. Then they are refined by using a membership function so as to obtain a video representation whose pixels are two-bit depth. Secondly, we design a non-uniform search point pattern and subsequently put forward an initial search point prediction algorithm based on two-bit depth presentation. Finally, centered on the predicted initial search point, the modified low-band-shift-based scalable video motion estimation (MLBSSME) is accepted to search motion vectors in a size-reduced window. Experimental results illustrate that the proposed algorithm can always achieve high motion estimation precision for video sequences with various characteristics. Our algorithm separately gains 0.41 dB and 1.43 dB higher average peak signal-to-noise ratio (PSNR) than the low-band-shift and band-to-band motion estimations, but 0.07 dB lower than full search (FS) in spatial domain. Nevertheless, the computational complexity of our algorithm is about 4.66% of FS, 4.62% of the low-band-shift motion estimation, and 22.70% of the band-to- band motion estimation.