An adaptive macroblock-mean difference based sorting scheme for fast normalized partial distortion search motion estimation

摘要

This work presents an efficient lossy partial distortion search (PDS) algorithm called adaptive mean difference based partial distortion search (AMDNPDS). The proposed AMDNPDS algorithm reduces computations by using a halfway-stop technique in the calculation of the macroblock (MB) distortion measure and applying a diagonal search pattern for stationary or quasi-stationary candidate MBs. For the matching point reduction, a MB is divided into 4 × 4 sub-MBs with each sub-MB sorted by subtracting the MB mean value. Therefore, the mean difference pixels are retrieved one at time to obtain the accumulated partial SAD used as a constraint for checking the validity of a candidate MB. The proposed scheme can accelerate the convergence speed and efficiently eliminate the impossible candidates earlier, resulting in substantial computation reduction. The experimental results show the proposed algorithm reduces the check pixels by about 11.02 times on average compared with the typical partial distortion search (PDS) when the motion MB size is 16 × 16 and the search range is ±15. Compared with other lossy PDS algorithm such as normalized PDS (NPDS), which achieved reductions of 1.82 times on average, reductions in computational complexity were achieved. In addition, the proposed algorithm achieved 59.78% of total motion estimation (ME) time saving compared to the NPDS algorithm and 58% total ME time in comparison to the prediction error prioritizing-based NPDS (PEP-NPDS) algorithm when using H.264/AVC JM 18.2 reference software according to different types of sequences, while maintaining a similar bit-rate without losing picture quality.