Nonlinear Scaling in Hierarchical Motion Estimation

The patent applies segmentation by dividing the motion estimation process into multiple hierarchical levels. Instead of performing a single exhaustive search over a large search window, the algorithm segments the search into coarse-to-fine stages: first identifying motion vectors at lower resolution levels with larger effective search windows, then refining these vectors at progressively higher resolutions. This segmentation allows the system to achieve coding gains comparable to large search windows while keeping encoding complexity manageable at each individual stage.

The patent implements the nested doll principle by creating a hierarchy of downsampled image layers nested within the full-resolution frame. Each layer is a scaled-down version of the previous layer, forming a nested structure where layer n+1 is embedded within the resolution space of layer n. This nested approach allows motion estimation to effectively access a large search window at coarser levels while using progressively smaller windows at finer levels, thereby achieving high motion estimation accuracy without requiring large on-chip memory buffers for all resolution levels simultaneously.

The patent applies dynamics by making the source block size adaptive rather than fixed. The algorithm dynamically adjusts the source block size based on local image characteristics, specifically the energy content of the region being encoded. In low-energy regions, larger source blocks are used to improve predictor accuracy by capturing more contextual information, while in high-energy regions, smaller blocks are used to maintain processing simplicity and adapt to local variations. This dynamic adaptation resolves the contradiction between processing simplicity and predictor accuracy.

The patent implements parameter changes by varying the downsampling factor across different hierarchical levels and adapting it to local image characteristics. Instead of uniformly downsampling by a factor of two at all levels, the algorithm adjusts the downsampling parameters dynamically. In regions with simple motion patterns, more aggressive downsampling is applied to maximize search speed, while in regions with complex motion or detailed structures, less aggressive downsampling is used to preserve motion estimation fidelity. This parameter adaptation allows the system to optimize the trade-off between search speed and fidelity on a region-by-region basis.