In video coding, accurate estimation of the motion information is very important for both motion-compensated prediction (MCP) and interpolation (MCI). Block-motion vector estimation by blockmatching assumes basically independent translational motion of small blocks. It often generates inaccurate or false motion vectors, that are locally optimized regardless of object motion. This is more so for natural images that have various spatial frequencies and diverse motion such as rotation and change of scale, in addition to the conventional translation. The use of inaccurately estimated block-motion vectors increases transmission bit cost of differential motion vectors, even though it makes some prediction gain in MCP. Furthermore, it degrades the interpolated image quality in MCI.
To solve this problem in block-motion vector estimation, this thesis proposes a new block-motion vector correction method which combines advantages of object-based motion estimation and motion vector postprocessing methods.
First, a new block-motion vector correction algorithm is proposed. The corrected motion vector is obtained from the motion description parameters of planar object, which estimated from the field of block-motion vectors. By depending on a weighting parameter and the measurement window size for the motion parameter estimation, the corrected motion is characterized either as global or local motion. It also provides motion vectors in pixel resolution as well as block resolution, that is very advantages in MCI applications. It is shown that the proposed algorithm estimates somewhat more accurate motion vectors than the conventional algorithms in real and artificially moved test images.
Second, a new hierarchical block-motion estimation algorithm based on the new motion vector correction is proposed. The first hierarchy measures global motion of longer range, and the second hierarchy provides local motion of shorter range. In each hierarchy, original block-motion vectors are obtained directly by blockmatching from given images. It is shown that the proposed algorithm produces accurate and homogeneous motion vectors, and hence gives improvements in prediction gain and transmission bit cost of motion vectors over the conventional algorithm.
Finally, effectiveness of the proposed method in MCI is investigated. It is shown that the MCI using this method requires smaller amount of frame memory and processing delay than the conventional method, while maintaining almost equal interpolated image quality.