Modified one-bit transform for motion estimation

Motion estimation using the one-bit transform (1BT) was proposed by Natarajan, Bhaskaran and Konstantinides (see ibid., vol.7, p.702-06, 1997) to achieve large computation reduction. However, it degrades the predicted image by almost 1 dB as compared with full search. We propose a modification to the 1BT by adding conditional local searches. Simulation results show that the proposed modification improves the peak signal-to-noise ratio (PSNR) significantly at the expense of slightly increased computational complexity. A variant of the proposed modification called the multiple-candidate two-step search (M2SSFS) is found to be particularly good for high quality, high bit rate video coding. In the MPEG-1 simulation, its PSNR is within 0.1 dB from that of full search at bit rates higher than 1 Mbit/s with a computation reduction factor of ten.