Threshold Mean Larger Ratio Motion Estimation in MPEG Encoding Using LNS

The Logarithmic Number System (LNS) offers good performance for some multimedia applications at low precisions. Motion estimation is a key part of the MPEG encoding system. Usually the motion estimation is performed by using a fixed-point Mean Absolute Difference (MAD) cost function for block matching. LNS Mean Larger Ratio (MLR) is another cost function which can produce comparable video encoding results as fixed-point MAD. This paper proposes an optimized Threshold MLR (TMLR) cost function, which uses a threshold value to keep only the larger LNS ratio of the pixel values with shorter word lengths. The encoding results show that the optimization makes significant area savings with little quality degradation. This paper also shows that the threshold approach can be applied to the fixed-point MAD. The Threshold MAD (TMAD) can achieve similar area savings as the TMLR does.