Spatially enhanced transform coding

What is this patent about?

’991 is related to the field of digital video compression and, more specifically, to the efficient encoding of prediction error signals. In standard hybrid video codecs, image blocks are predicted using temporal or spatial methods, and the resulting residual—the difference between the original and the predicted pixels—is typically compressed using frequency-domain transforms like the Discrete Cosine Transform. While these transforms excel at packing energy for highly correlated data, they often struggle with localized irregularities or high-frequency noise, leading to coding inefficiencies.

The underlying idea behind ’991 is to treat the prediction error signal as a composite of two distinct types of data: well-correlated components suitable for frequency transforms and localized outlier values that are better handled in the spatial domain. Rather than forcing a transform to account for sharp, isolated pixel differences—which would scatter energy across many coefficients—the invention isolates these outliers. By coding the bulk of the signal through a transform and the remaining localized errors through direct spatial quantization, the system achieves a higher coding gain than either method could provide alone.

The claims of ’991 focus on a hybrid coding architecture that generates a prediction error signal by joining a transform-coded representation with a spatially-coded representation of the same difference signal. The independent claims cover both the encoding process—where a difference signal is decomposed into these two components—and the corresponding decoding process, which reconstructs the image block by summing the decoded transform information, the decoded spatial samples, and the original prediction.

In a practical implementation, the encoder identifies specific pixels where the prediction error is significantly higher than the surrounding neighborhood. These outliers are temporarily replaced with interpolated values to create a smooth, modified signal that the transform can compress highly efficiently. The difference between the original outlier and its smoothed version is then captured as a spatial sample, ensuring that the fine details or sharp edges are not lost or blurred by the quantization of frequency coefficients.

This approach differs from prior solutions by moving away from a strict 'either-or' choice between transform and spatial domains for a given block. Instead of selecting one mode for an entire macroblock, the invention allows concurrent coding of both types within the same data unit. This dual-path reconstruction enables the codec to maintain high fidelity for textures and edges while leveraging the energy compaction of transforms for the rest of the signal, effectively reducing the total bitrate required for high-quality video.