A Novel Prediction-Based Asymmetric Fast Search Algorithm for Video Compression

Abstract

The successive frames of a video sequence are highly correlated. Therefore, by reducing temporal redundancy a high compression ratio can be achieved. Motion estimation, which reduces temporal redundancy, plays an important role in video coding systems such as H.26x and MPEG-x. The block-matching algorithm (BMA) is the most popular and widely used algorithm for motion estimation due to its simplicity and reasonable performance. In BMA, an image frame is divided into non-overlapping rectangular blocks with equal or variable block sizes. The pixels in each block are assumed to have the same motion. The motion vector (MV) of a block is estimated by searching for its best match within a search window in the previous frame. The distortion between the current block and each searching block is employed as a matching criterion. The resulting MV is used to generate a motioncompensated prediction block. The motion-compensated prediction difference blocks (called residue blocks) and the MVs are encoded and then sent to the decoder. Among the various BMAs, the full search (FS) is global optimal and most straightforward algorithm because it searches the entire search window for the best matching block. However, its only drawback is a heavy computational load. To address this drawback, many fast BMAs have been proposed in the literature since 1981, such as the three-step search (TSS) (Koga et al., 1981), cross search (Ghanbari, 1990), new three-step search (NTSS) (Li et al., 1994), block-based gradient descent search (Liu & Feig,1996), four-step search (Po & Ma, 1996), diamond search (DS) (Zhu & Ma, 2000), hexagon-based search (HEXBS) (Zhu et al., 2002), and the cross-diamond search (CDS) (Cheung & Po, 2002), etc. The primary assumption of most fast BMAs is that the block distortion is monotonic over the search range, implying that the distortion decreases monotonically as the search position moves toward the minimum distortion point. Therefore, the best match point can be found by following the distortion trend without checking all search points in the search window. Consequently, these fast BMAs use various search patterns to reduce the number of search points, thereby speeding up the search. In addition, some fast BMAs (Liu & Zaccarin, 1993; Yu et al., 2001; Bierling, 1988) speed up the search by sub-sampling the block pixels on distortion computation and/or by sub-sampling the search points in the search window. Furthermore, since there is a high correlation between a current block and its adjacent blocks in spatial and/or temporal domains, the