Method for sub-pixel value interpolation

What is this patent about?

’273 is related to the field of digital video coding and decoding, specifically focusing on motion-compensated temporal prediction. In modern video compression, the accuracy of motion vectors is critical for reducing prediction error; however, real-world motion often occurs at a finer resolution than the original pixel grid. The patent addresses the computational and memory challenges associated with interpolating sub-pixel values at fractional locations, such as half, quarter, or eighth-pixel resolutions, to improve the efficiency of video codecs.

The underlying idea behind ’273 is a multi-stage interpolation strategy that balances computational precision with hardware efficiency by utilizing a specific sequence of filtering and averaging. The core insight is to derive primary sub-pixel values using high-order filters (such as 6-tap or 8-tap filters) to maintain accuracy, and then generate finer fractional values through simpler linear or diagonal averages. By calculating certain sub-pixels diagonally from previously interpolated values, the invention avoids the cumulative rounding errors and high memory overhead found in prior methods that rely on exhaustive multi-step horizontal and vertical filtering.

The claims of ’273 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 a process where sub-pixels at odd-numbered fractional coordinates (the most granular level) are interpolated using a weighted average of either a nearest-neighbor pixel and a central half-pixel, or a pair of diagonally-opposed sub-pixels. This diagonal approach is supported by first calculating intermediate sub-pixels at even fractional coordinates using weighted sums of pixels in the primary rows and columns.

In practice, the invention operates by first applying a Kth-order filter to original pixels to generate half-resolution values, which are stored as intermediate results. When a motion vector points to a quarter-pixel or eighth-pixel location, the system performs diagonal interpolation to reach the final value. This differs from prior approaches like TML5, which suffered from precision loss due to multiple stages of truncation, and TML6, which required high-precision 32-bit arithmetic for all calculations. The '273 method allows for high-quality prediction while keeping the arithmetic requirements low enough for standard CPUs and low-cost ASICs.

The differentiation of this invention lies in its flexibility and reduced computational burden during the decoding process. By allowing a choice between horizontal and vertical filtering to reach the same intermediate sub-pixel, the decoder can select the path that requires the fewest operations for a specific motion vector. This on-demand interpolation strategy significantly reduces the number of processor cycles and the amount of silicon area needed for memory, making it particularly suitable for mobile terminals and real-time videotelephony where power and hardware resources are constrained.