Method for sub-pixel value interpolation

What is this patent about?

’599 is related to the field of digital video coding and, more specifically, to methods for sub-pixel value interpolation used in motion-compensated temporal prediction. In modern video compression, motion vectors often point to fractional pixel locations to more accurately model the movement of objects between frames. This requires the system to estimate the values of these sub-pixels based on the known values of original pixels in a reference frame, a process that traditionally balances the trade-off between computational complexity, memory consumption, and interpolation precision.

The underlying idea behind ’599 is a multi-stage interpolation strategy that reduces computational dependency and maintains high precision without requiring excessive memory. Instead of calculating all fractional values through a single complex filter or relying on a chain of previously rounded sub-pixels, the invention utilizes a specific hierarchy. It first generates half-resolution values using high-order filters and then derives finer quarter- or eighth-resolution values through diagonal linear interpolation. This approach avoids the accumulation of rounding errors found in prior methods while ensuring that the decoder can calculate only the specific sub-pixel needed with minimal overhead.

The claims of ’599 focus on a method for determining sub-pixel values within a rectangular region defined by four corner pixels using a specific coordinate-based logic. The independent claims cover the interpolation of sub-pixels at locations with odd coordinate values (such as 1/4 or 3/4 positions) by taking a weighted average of diagonally-opposed elements, such as a corner pixel and a central half-pixel. This diagonal approach is supported by the preliminary calculation of half-unit values using weighted sums of pixels in rows and columns, providing a structured framework for reaching higher levels of fractional precision.

In practice, the invention works by first applying a high-order filter (such as a 6-tap or 8-tap filter) to original pixels to create intermediate half-pixel values. For the most complex fractional positions—those residing at odd horizontal and vertical offsets—the system performs a simple average of two previously calculated diagonal points. This diagonal symmetry is a key differentiator from prior approaches like TML5, which relied on a more rigid horizontal-then-vertical sequence that often resulted in lower precision due to multiple stages of truncation and clipping.

By decoupling the calculation of fine-resolution sub-pixels from a strictly linear row-column processing order, the invention achieves a significant reduction in the number of required filtering operations. It allows the decoder to be more flexible, choosing the most efficient path to a specific sub-pixel value based on its location. Furthermore, by using intermediate high-precision values for the initial filtering stage and simple linear averages for the final stage, the method provides a high-quality prediction frame that is both computationally lean for mobile devices and memory-efficient for high-performance encoders.