Pipelined macroblock processing to reduce internal buffer size of motion estimation in multimedia SoCs

Abstract

A multimedia SoC often requires a large internal buffer, because it must store the whole search window to reduce the huge I/O bandwidth of motion estimation. However, the silicon area of the internal buffer increases tremendously as the search range becomes larger. This paper proposes a new method that greatly reduces the internal buffer size of a multimedia SoC while the computational cost, I/O bandwidth, and image quality do not change. In the proposed method, only the overlapped parts of search windows for consecutive macroblocks are stored in the internal buffer. The proposed method reduces the internal buffer size to 1/5.0 and 1/8.8 when the search range is +/-64 X +/-64 and +/-128 X +/-128, respectively.